{"id":24819,"date":"2021-03-24T18:35:00","date_gmt":"2021-03-24T10:35:00","guid":{"rendered":"https:\/\/swarma.org\/?p=24819"},"modified":"2021-03-24T18:35:00","modified_gmt":"2021-03-24T10:35:00","slug":"%e9%95%bf%e6%96%87%e7%bb%bc%e8%bf%b0%ef%bc%9a%e5%9b%be%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e7%9a%84%e5%8f%af%e8%a7%a3%e9%87%8a%e6%80%a7","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=24819","title":{"rendered":"\u957f\u6587\u7efc\u8ff0\uff1a\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u53ef\u89e3\u91ca\u6027"},"content":{"rendered":"
\n
\n
\n
\n

<\/p>\n<\/section>\n

\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n

\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u53ef\u89e3\u91ca\u6027\u662f\u76ee\u524d\u6bd4\u8f83\u503c\u5f97\u63a2\u7d22\u7684\u65b9\u5411\uff0c\u4eca\u5929\u89e3\u8bfb\u76842021\u6700\u65b0\u7efc\u8ff0\uff0c\u5176\u9488\u5bf9\u8fd1\u671f\u63d0\u51fa\u7684 GNN \u89e3\u91ca\u6280\u672f\u8fdb\u884c\u4e86\u7cfb\u7edf\u7684\u603b\u7ed3\u548c\u5206\u6790\uff0c\u5f52\u7eb3\u5bf9\u6bd4\u4e86\u8be5\u95ee\u9898\u7684\u89e3\u51b3\u601d\u8def\u3002\u4f5c\u8005\u8fd8\u4e3aGNN\u89e3\u91ca\u6027\u95ee\u9898\u63d0\u4f9b\u4e86\u6807\u51c6\u7684\u56fe\u6570\u636e\u96c6\u548c\u8bc4\u4f30\u6307\u6807\uff0c\u5c06\u662f\u8fd9\u4e00\u65b9\u5411\u975e\u5e38\u503c\u5f97\u53c2\u8003\u7684\u4e00\u7bc7\u6587\u7ae0\u3002<\/span><\/strong><\/p>\n


<\/strong><\/p>\n

\u96c6\u667a\u4ff1\u4e50\u90e8\u8054\u5408\u667a\u6e90\u793e\u533a\uff0c\u4ee5\u56e0\u679c\u79d1\u5b66\u548cCausal AI\u4e3a\u4e3b\u9898\u4e3e\u529e\u7cfb\u5217\u8bfb\u4e66\u4f1a<\/a>\uff0c\u7cbe\u8bfb\u57fa\u7840\u6559\u6750\u3001\u7814\u8bfb\u91cd\u8981\u8bba\u6587\uff0c\u63a2\u8ba8\u5982\u4f55\u501f\u52a9\u56e0\u679c\u79d1\u5b66\u6784\u5efa\u53ef\u89e3\u91ca\u7684\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u3002\u8be6\u60c5\u89c1\u6587\u672b\u3002<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
<\/p>\n

\u56fe\u5b50<\/span> <\/span>| \u4f5c\u8005<\/span><\/strong><\/strong><\/h3>\n

<\/span><\/p>\n

\u56fe\u4e0e\u63a8\u8350<\/span> | \u6765\u6e90<\/strong><\/span><\/p>\n


<\/p>\n<\/section>\n


<\/section>\n
\n

<\/figcaption><\/figure>\n<\/section>\n

\n
\n
\n
\n

\u8bba\u6587\u9898\u76ee\uff1a<\/span><\/p>\n

Explainability in Graph Neural Networks: A Taxonomic Survey<\/span><\/section>\n
\u8bba\u6587\u5730\u5740\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2012.15445<\/section>\n<\/section>\n<\/section>\n<\/blockquote>\n<\/h1>\n


<\/p>\n

<\/p>\n

\u626b\u7801\u4e8c\u7ef4\u7801\u6216\u70b9\u51fb\u6587\u672b\u201c\u9605\u8bfb\u539f\u6587\u201d\u5373\u53ef\u4e0b\u8f7dPDF\u6587\u4ef6<\/span><\/strong><\/p>\n

2\u4e07\u5b57\u957f\u6587\uff0c\u5efa\u8bae\u6536\u85cf\u6162\u6162\u770b~<\/span><\/strong>
<\/section>\n
\u76ee \u5f55<\/span><\/strong><\/span><\/section>\n
    \n
  • \n
    0.Abstract<\/span><\/section>\n<\/li>\n
  • \n
    1. Introduction<\/span><\/section>\n<\/li>\n
  • \n
    2. \u603b\u4f53\u6846\u67b6<\/span><\/section>\n<\/li>\n
  • \n
    3.\u65b9\u6cd5\u4ecb\u7ecd<\/span><\/section>\n<\/li>\n
      \n
    • \n
      3.1 \u57fa\u4e8e\u68af\u5ea6\/\u7279\u5f81\u7684\u65b9\u6cd5\uff08Gradients\/Features-Based Methods\uff09<\/span><\/section>\n<\/li>\n
    • \n
      3.2 \u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\uff08Perturbation-Based Methods\uff09<\/span><\/section>\n<\/li>\n
    • \n
      3.3 \u57fa\u4e8e\u4ee3\u7406\u7684\u65b9\u6cd5\uff08Surrogate Methods\uff09<\/span><\/section>\n<\/li>\n
    • \n
      3.4 \u5206\u89e3\u65b9\u6cd5\uff08Decomposition Methods\uff09<\/span><\/section>\n<\/li>\n<\/ul>\n
    • \n
      4. \u6a21\u578b\u7ea7\u65b9\u6cd5<\/span><\/section>\n<\/li>\n
    • \n
      5. \u8bc4\u4f30\u6a21\u578b<\/span><\/section>\n<\/li>\n
        \n
      • \n
        5.1. Datasets<\/span><\/section>\n<\/li>\n
      • \n
        5.1.1. Synthetic data<\/span><\/section>\n<\/li>\n
      • \n
        5.1.2 Sentiment graph data<\/span><\/section>\n<\/li>\n
      • \n
        5.1.3 Molecule data<\/span><\/section>\n<\/li>\n
      • \n
        5.2 Evaluation Metrics<\/span><\/section>\n<\/li>\n
      • \n
        5.2.1 Fidelity\/Infidelity<\/span><\/section>\n<\/li>\n
      • \n
        5.2.2 Sparsity<\/span><\/section>\n<\/li>\n
      • \n
        5.2.3 Stability<\/span><\/section>\n<\/li>\n
      • \n
        5.2.4 Accuracy<\/span><\/section>\n<\/li>\n<\/ul>\n
      • \n
        6. Conclusion<\/span><\/section>\n<\/li>\n
      • \n
        \u53c2\u8003\u6587\u732e<\/span><\/section>\n<\/li>\n<\/ul>\n
        <\/p>\n
        \n


        <\/p>\n

         <\/span><\/p>\n<\/section>\n

        <\/span><\/section>\n

        \n
        \n
        \n
        \n
        \n

        <\/section>\n

        <\/section>\n
        \n
        <\/svg><\/section>\n<\/section>\n<\/section>\n
        \n

        0. Abstract<\/strong><\/strong><\/p>\n<\/section>\n

        \n

        <\/section>\n

        <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
        \u8fd1\u5e74\u6765\uff0c\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u7814\u7a76\u5728\u56fe\u50cf\u548c\u6587\u672c\u9886\u57df\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\u3002\u7136\u800c\uff0c\u5728\u56fe\u6570\u636e\u9886\u57df\uff0c\u65e2\u6ca1\u6709\u9488\u5bf9GNN\u53ef\u89e3\u91ca\u6027\u7684\u7edf\u4e00\u5904\u7406\u65b9\u6cd5\uff0c\u4e5f\u4e0d\u5b58\u5728\u6807\u51c6\u7684 benchmark \u6570\u636e\u96c6\u548c\u8bc4\u4f30\u51c6\u5219\u3002\u5728\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u4f5c\u8005\u5bf9\u76ee\u524d\u7684GNN\u89e3\u91ca\u6280\u672f\u4ece\u7edf\u4e00\u548c\u5206\u7c7b\u7684\u89d2\u5ea6\u8fdb\u884c\u4e86\u603b\u7ed3\uff0c\u9610\u660e\u4e86\u73b0\u6709\u65b9\u6cd5\u7684\u5171\u6027\u548c\u5dee\u5f02\uff0c\u5e76\u4e3a\u8fdb\u4e00\u6b65\u7684\u65b9\u6cd5\u53d1\u5c55\u5960\u5b9a\u4e86\u57fa\u7840\u3002\u6b64\u5916\uff0c\u4f5c\u8005\u4e13\u95e8\u4e3aGNN\u89e3\u91ca\u6280\u672f\u751f\u6210\u4e86 benchmark \u56fe\u6570\u636e\u96c6\uff0c\u5e76\u603b\u7ed3\u4e86\u5f53\u524d\u7528\u4e8e\u8bc4\u4f30GNN\u89e3\u91ca\u6280\u672f\u7684\u6570\u636e\u96c6\u548c\u8bc4\u4f30\u65b9\u6cd5\u3002<\/span><\/section>\n
        <\/p>\n
        \n

         <\/span><\/p>\n<\/section>\n

        <\/span><\/section>\n

        \n
        \n
        \n
        \n
        \n

        <\/section>\n

        <\/section>\n
        \n
        <\/svg><\/section>\n<\/section>\n<\/section>\n
        \n

        1. Introduction<\/strong><\/strong><\/p>\n<\/section>\n

        \n

        <\/section>\n

        <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
        \u89e3\u91ca\u9ed1\u7bb1\u6a21\u578b\u662f\u5341\u5206\u5fc5\u8981\u7684\uff1a\u5982\u679c\u6ca1\u6709\u5bf9\u9884\u6d4b\u80cc\u540e\u7684\u5e95\u5c42\u673a\u5236\u8fdb\u884c\u63a8\u7406\uff0c\u6df1\u5c42\u6a21\u578b\u5c31\u65e0\u6cd5\u5f97\u5230\u5b8c\u5168\u4fe1\u4efb\uff0c\u8fd9\u5c31\u963b\u788d\u4e86\u6df1\u5ea6\u6a21\u578b\u5728\u4e0e\u516c\u5e73\u6027\u3001\u9690\u79c1\u6027\u548c\u5b89\u5168\u6027\u6709\u5173\u7684\u5173\u952e\u5e94\u7528\u7a0b\u5e8f\u4e2d\u4f7f\u7528\u3002\u4e3a\u4e86\u5b89\u5168\u3001\u53ef\u4fe1\u5730\u90e8\u7f72\u6df1\u5ea6\u6a21\u578b\uff0c\u9700\u8981<\/span>\u540c\u65f6\u63d0\u4f9b\u51c6\u786e\u7684\u9884\u6d4b\u548c\u4eba\u7c7b\u80fd\u9886\u4f1a\u7684\u89e3\u91ca<\/span><\/strong>\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u8de8\u5b66\u79d1\u9886\u57df\u7684\u7528\u6237\u3002<\/span><\/section>\n
        \u6df1\u5c42\u6a21\u578b\u7684\u89e3\u91ca\u6280\u672f\u901a\u5e38\u4ece\u7814\u7a76\u5176\u9884\u6d4b\u80cc\u540e\u7684\u6f5c\u5728\u5173\u7cfb\u7740\u624b\uff0c\u89e3\u91ca\u6280\u672f\u5927\u81f4\u53ef\u5206\u4e3a\u4e24\u7c7b\uff1a<\/span><\/section>\n
        1\uff09input-dependent explanations\uff08\u4f9d\u8d56\u8f93\u5165\u7684\u89e3\u91ca\u65b9\u6cd5\uff09<\/span><\/h5>\n
        \u8be5\u7c7b\u65b9\u6cd5\u4ece\u7279\u5f81\u7684\u89d2\u5ea6\u51fa\u53d1\uff0c\u63d0\u4f9b\u4e0e\u8f93\u5165\u76f8\u5173\u7684\u89e3\u91ca\uff0c\u4f8b\u5982\u7814\u7a76\u8f93\u5165\u7279\u5f81\u7684\u91cd\u8981\u6027\u5f97\u5206\uff0c\u6216\u5bf9\u6df1\u5c42\u6a21\u578b\u7684\u4e00\u822c\u884c\u4e3a\u6709\u9ad8\u6c34\u5e73\u7684\u7406\u89e3\u3002\u8bba\u6587 [10]\uff0c[11]\uff0c[18]\u901a\u8fc7\u7814\u7a76\u68af\u5ea6\u6216\u6743\u91cd\uff0c\u5206\u6790\u9884\u6d4b\u7ed3\u679c\u76f8\u5bf9\u4e8e\u8f93\u5165\u7279\u5f81\u7684\u654f\u611f\u7a0b\u5ea6\u3002\u8bba\u6587 [12]\uff0c[13]\uff0c[19] \u901a\u8fc7\u5c06\u9690\u85cf\u7279\u5f81\u6620\u5c04\u5230\u8f93\u5165\u7a7a\u95f4\uff0c\u4ece\u800c\u7a81\u51fa\u91cd\u8981\u7684\u8f93\u5165\u7279\u5f81\u3002[14] \u901a\u8fc7\u906e\u853d\u4e0d\u540c\u7684\u8f93\u5165\u7279\u5f81\uff0c\u89c2\u5bdf\u9884\u6d4b\u7684\u53d8\u5316\uff0c\u4ee5\u8bc6\u522b\u91cd\u8981\u7684\u7279\u5f81\u3002<\/span><\/section>\n
        2\uff09input-independent explanations\uff08\u72ec\u7acb\u4e8e\u8f93\u5165\u7684\u89e3\u91ca\u65b9\u6cd5\uff09<\/span><\/h5>\n
        \u4e0e\u4f9d\u8d56\u7279\u5f81\u7684\u89e3\u91ca\u65b9\u6cd5\u4e0d\u540c\uff0c\u8be5\u7c7b\u65b9\u6cd5\u4ece\u6a21\u578b\u89d2\u5ea6\u51fa\u53d1\uff0c\u63d0\u4f9b\u72ec\u7acb\u4e8e\u8f93\u5165\u7684\u89e3\u91ca\uff0c\u4f8b\u5982\u7814\u7a76\u8f93\u5165 patterns\uff0c\u4f7f\u67d0\u7c7b\u7684\u9884\u6d4b\u5f97\u5206\u6700\u5927\u5316\u3002\u8bba\u6587[17]\uff0c[22]\u901a\u8fc7\u63a2\u7a76\u9690\u85cf\u795e\u7ecf\u5143\u7684\u542b\u4e49\uff0c\u8fdb\u800c\u7406\u89e3\u7684\u6574\u4e2a\u9884\u6d4b\u8fc7\u7a0b\u3002\u8bba\u6587[23]\uff0c[24]\uff0c[25]\uff0c[26] \u5bf9\u8fd1\u671f\u7684\u65b9\u6cd5\u8fdb\u884c\u4e86\u8f83\u4e3a\u7cfb\u7edf\u7684\u8bc4\u4ef7\u548c\u5206\u7c7b\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u7814\u7a76\u53ea\u5173\u6ce8\u56fe\u50cf\u548c\u6587\u672c\u9886\u57df\u7684\u89e3\u91ca\u65b9\u6cd5\uff0c\u5ffd\u7565\u4e86\u6df1\u5ea6\u56fe\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u3002<\/span><\/section>\n

        GNN \u7684\u53ef\u89e3\u91ca\u6027<\/span><\/h4>\n
        \u4e0e\u56fe\u50cf\u548c\u6587\u672c\u9886\u57df\u76f8\u6bd4\uff0c\u5bf9\u56fe\u6a21\u578b\u89e3\u91ca\u6027\u7684\u7814\u7a76\u8f83\u5c11\uff0c\u7136\u800c\u8fd9\u662f\u7406\u89e3\u6df1\u5ea6\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u5173\u952e\u3002\u8fd1\u5e74\u6765\uff0c\u4eba\u4eec\u63d0\u51fa\u4e86\u51e0\u79cd\u89e3\u91ca GNN \u9884\u6d4b\u7684\u65b9\u6cd5\uff0c\u5982XGNN[41]\u3001gnexplainer[42]\u3001PGExplainer[43]\u7b49\u3002\u8fd9\u4e9b\u65b9\u6cd5\u662f\u4ece\u4e0d\u540c\u7684\u89d2\u5ea6\u63d0\u4f9b\u4e86\u4e0d\u540c\u5c42\u6b21\u7684\u89e3\u91ca\u3002\u4f46\u81f3\u4eca\u4ecd\u7136**\u7f3a\u4e4f\u6807\u51c6\u7684\u6570\u636e\u96c6\u548c\u5ea6\u91cf\u6765\u8bc4\u4f30\u89e3\u91ca\u7ed3\u679c\u3002**\u56e0\u6b64\uff0c\u9700\u8981\u5bf9GNN\u89e3\u91ca\u6280\u672f\u548c\u5176\u8bc4\u4f30\u65b9\u6cd5\u8fdb\u884c\u7cfb\u7edf\u7684\u7814\u7a76\u3002<\/span><\/section>\n

        \u672c\u6587<\/span><\/h4>\n
        \u672c\u7814\u7a76\u63d0\u4f9b\u4e86\u5bf9\u4e0d\u540cGNN\u89e3\u91ca\u6280\u672f\u7684\u7cfb\u7edf\u7814\u7a76\uff0c\u76ee\u7684\u5bf9\u4e0d\u540c\u65b9\u6cd5\u8fdb\u884c\u76f4\u89c2\u548c\u9ad8\u6c34\u5e73\u7684\u89e3\u91ca\uff0c\u8bba\u6587\u8d21\u732e\u5982\u4e0b\uff1a<\/span><\/section>\n
          \n
        • \n
          \u5bf9\u73b0\u6709\u7684\u6df1\u5ea6\u56fe\u6a21\u578b\u7684\u89e3\u91ca\u6280\u672f\u8fdb\u884c\u4e86\u7cfb\u7edf\u548c\u5168\u9762\u7684\u56de\u987e\u3002<\/span><\/section>\n<\/li>\n
        • \n
          \u63d0\u51fa\u4e86\u73b0\u6709GNN\u89e3\u91ca\u6280\u672f\u7684\u65b0\u578b\u5206\u7c7b\u6846\u67b6\uff0c\u603b\u7ed3\u4e86\u6bcf\u4e2a\u7c7b\u522b\u7684\u5173\u952e\u601d\u60f3\uff0c\u5e76\u8fdb\u884c\u4e86\u6df1\u523b\u7684\u5206\u6790\u3002<\/span><\/section>\n<\/li>\n
        • \n
          \u8be6\u7ec6\u4ecb\u7ecd\u4e86\u6bcf\u79cdGNN\u89e3\u91ca\u65b9\u6cd5\uff0c\u5305\u62ec\u5176\u65b9\u6cd5\u8bba\u3001\u4f18\u52bf\u3001\u7f3a\u70b9\uff0c\u4e0e\u5176\u4ed6\u65b9\u6cd5\u7684\u533a\u522b\u3002<\/span><\/section>\n<\/li>\n
        • \n
          \u603b\u7ed3\u4e86GNN\u89e3\u91ca\u4efb\u52a1\u4e2d\u5e38\u7528\u7684\u6570\u636e\u96c6\u548c\u8bc4\u4ef7\u6307\u6807\uff0c\u8ba8\u8bba\u4e86\u5b83\u4eec\u7684\u5c40\u9650\u6027\uff0c\u5e76\u63d0\u51fa\u4e86\u51e0\u70b9\u5efa\u8bae\u3002<\/span><\/section>\n<\/li>\n
        • \n
          \u901a\u8fc7\u5c06\u53e5\u5b50\u8f6c\u6362\u4e3a\u56fe\uff0c\u9488\u5bf9\u6587\u672c\u9886\u57df\u6784\u5efa\u4e86\u4e09\u4e2a\u4eba\u7c7b\u53ef\u7406\u89e3\u7684\u6570\u636e\u96c6\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u5373\u5c06\u516c\u5f00\uff0c\u53ef\u4ee5\u76f4\u63a5\u7528\u4e8eGNN\u89e3\u91ca\u4efb\u52a1\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

          \u540d\u8bcd\u89e3\u91ca\uff1aExplainability versus Interpretability<\/span><\/h4>\n
          \u5728\u4e00\u4e9b\u7814\u7a76\u4e2d\uff0c\u201cexplainability\u201d \u548c \u201cinterpretability\u201d\u88ab\u4ea4\u66ff\u4f7f\u7528\u3002\u672c\u6587\u4f5c\u8005\u8ba4\u4e3a\u8fd9\u4e24\u4e2a\u672f\u8bed\u5e94\u8be5\u88ab\u533a\u5206\u5f00\u6765\uff0c\u9075\u5faa\u8bba\u6587[44]\u6765\u533a\u5206\u8fd9\u4e24\u4e2a\u672f\u8bed\u3002\u5982\u679c\u4e00\u4e2a<\/span>\u6a21\u578b\u672c\u8eab<\/span><\/strong>\u80fd\u591f\u5bf9\u5176\u9884\u6d4b\u63d0\u4f9b\u4eba\u7c7b\u53ef\u7406\u89e3\u7684\u89e3\u91ca\uff0c\u5219\u8ba4\u4e3a\u8fd9\u4e2a\u6a21\u578b\u662f “interpretable”\u3002\u6ce8\u610f\uff0c\u8fd9\u6837\u7684\u6a21\u578b\u5728\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u4e0d\u518d\u662f\u4e00\u4e2a\u9ed1\u76d2\u5b50\u3002\u4f8b\u5982\uff0c\u4e00\u4e2a\u51b3\u7b56\u6811\u6a21\u578b\u5c31\u662f\u4e00\u4e2a “interpretable\u201c\u7684\u6a21\u578b\u3002\u540c\u65f6\uff0c”explainable “\u6a21\u578b\u610f\u5473\u7740\u8be5\u6a21\u578b\u4ecd\u7136\u662f\u4e00\u4e2a\u9ed1\u76d2\u5b50\uff0c\u5176\u9884\u6d4b\u6709\u53ef\u80fd\u88ab\u4e00\u4e9b<\/span>\u4e8b\u540e\u89e3\u91ca<\/span><\/strong>\u6280\u672f\u6240\u7406\u89e3\u3002<\/span><\/section>\n
          <\/p>\n
          \n

           <\/span><\/p>\n<\/section>\n

          <\/span><\/section>\n

          \n
          \n
          \n
          \n
          \n

          <\/section>\n

          <\/section>\n
          \n
          <\/svg><\/section>\n<\/section>\n<\/section>\n
          \n

          2. \u603b\u4f53\u6846\u67b6<\/strong><\/strong><\/p>\n<\/section>\n

          \n

          <\/section>\n

          <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
          \u76ee\u524d\u5b58\u5728\u4e00\u7cfb\u5217\u9488\u5bf9\u6df1\u5ea6\u56fe\u6a21\u578b\u89e3\u91ca\u6027\u95ee\u9898\u7684\u5de5\u4f5c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u5173\u6ce8\u56fe\u6a21\u578b\u7684\u4e0d\u540c\u65b9\u9762\uff0c\u5e76\u63d0\u4f9b\u4e0d\u540c\u7684\u89c2\u70b9\u6765\u7406\u89e3\u8fd9\u4e9b\u6a21\u578b\u3002\u5b83\u4eec\u4e00\u822c\u90fd\u4f1a\u4ece\u51e0\u4e2a\u95ee\u9898\u51fa\u53d1\u5b9e\u73b0\u5bf9\u56fe\u6a21\u578b\u7684\u89e3\u91ca\uff1a<\/span>\u54ea\u4e9b\u8f93\u5165\u8fb9\u66f4\u91cd\u8981\uff1f\u54ea\u4e9b\u8f93\u5165\u8282\u70b9\u66f4\u91cd\u8981\uff1f<\/span><\/strong> \u54ea\u4e9b\u8282\u70b9\u7279\u5f81\u66f4\u91cd\u8981\uff1f\u4ec0\u4e48\u6837\u7684\u56fe\u6a21\u5f0f\u4f1a\u6700\u5927\u9650\u5ea6\u5730\u9884\u6d4b\u67d0\u4e2a\u7c7b\uff1f\u4e3a\u4e86\u66f4\u597d\u5730\u7406\u89e3\u8fd9\u4e9b\u65b9\u6cd5\uff0c\u672c\u6587\u4e3aGNNs\u7684\u4e0d\u540c\u89e3\u91ca\u6280\u672f\u63d0\u4f9b\u4e86\u5206\u7c7b\u6846\u67b6\uff0c\u7ed3\u6784\u5982\u56fe1\u6240\u793a\u3002\u6839\u636e\u63d0\u4f9b\u4ec0\u4e48\u7c7b\u578b\u7684\u89e3\u91ca\uff0c\u89e3\u91ca\u6027\u6280\u672f\u88ab\u5206\u4e3a\u4e24\u5927\u7c7b\uff1a\u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u548c\u6a21\u578b\u7ea7\u65b9\u6cd5\u3002<\/span>\u672c\u6587\u63a5\u4e0b\u6765\u7684\u90e8\u5206\u5c06\u9488\u5bf9\u56fe1\u7684\u5404\u4e2a\u5206\u652f\u5c55\u5f00\u8bb2\u89e3\uff0c\u5e76\u4f5c\u51fa\u5bf9\u6bd4\u3002<\/span><\/strong><\/section>\n
          <\/div>\n
          \u56fe1 GNN \u89e3\u91ca\u6027\u65b9\u6cd5\u5206\u7c7b\u6846\u67b6<\/span><\/section>\n

          1\uff09\u5b9e\u4f8b\u7ea7\u65b9\u6cd5<\/span><\/h4>\n
          \u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u4e0e\u7279\u5f81\u5de5\u7a0b\u7684\u601d\u60f3\u6709\u4e9b\u7c7b\u4f3c\uff0c\u65e8\u5728\u627e\u5230\u8f93\u5165\u6570\u636e\u4e2d\u6700\u80fd\u591f\u5f71\u54cd\u9884\u6d4b\u7ed3\u679c\u7684\u90e8\u5206\u7279\u5f81\uff0c\u4e3a\u6bcf\u4e2a\u8f93\u5165\u56fe\u63d0\u4f9b input-dependent \u7684\u89e3\u91ca\u3002\u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u56fe\uff0c<\/span>\u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u65e8\u5728\u63a2\u7a76\u5f71\u54cd\u6a21\u578b\u9884\u6d4b\u7684\u91cd\u8981\u7279\u5f81\u5b9e\u73b0\u5bf9\u6df1\u5ea6\u6a21\u578b\u7684\u89e3\u91ca<\/span><\/strong>\u3002\u6839\u636e\u7279\u5f81\u91cd\u8981\u6027\u5206\u6570\u7684\u83b7\u5f97\u65b9\u5f0f\uff0c\u4f5c\u8005\u5c06\u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u5206\u4e3a\u56db\u4e2a\u4e0d\u540c\u7684\u5206\u652f\uff1a<\/span><\/section>\n
            \n
          • \n
            \u57fa\u4e8e\u68af\u5ea6\/\u7279\u5f81\u7684\u65b9\u6cd5[49]\uff0c[50]\uff0c\u91c7\u7528\u68af\u5ea6\u6216\u7279\u5f81\u503c\u6765\u8868\u793a\u4e0d\u540c\u8f93\u5165\u7279\u5f81\u7684\u91cd\u8981\u7a0b\u5ea6\u3002<\/span><\/section>\n<\/li>\n
          • \n
            \u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5[42]\uff0c[43]\uff0c[51]\uff0c[52]\uff0c[53]\uff0c\u76d1\u6d4b\u5728\u4e0d\u540c\u8f93\u5165\u6270\u52a8\u4e0b\u9884\u6d4b\u503c\u7684\u53d8\u5316\uff0c\u4ece\u800c\u5b66\u4e60\u8f93\u5165\u7279\u5f81\u7684\u91cd\u8981\u6027\u5206\u6570\u3002<\/span><\/section>\n<\/li>\n
          • \n
            \u57fa\u4e8e\u5206\u89e3\u7684\u65b9\u6cd5[49]\uff0c[50]\uff0c[54]\uff0c[55]\uff0c\u9996\u5148\u5c06\u9884\u6d4b\u5206\u6570\uff0c\u5982\u9884\u6d4b\u6982\u7387\uff0c\u5206\u89e3\u5230\u6700\u540e\u4e00\u4e2a\u9690\u85cf\u5c42\u7684\u795e\u7ecf\u5143\u3002\u7136\u540e\u5c06\u8fd9\u6837\u7684\u5206\u6570\u9010\u5c42\u53cd\u5411\u4f20\u64ad\uff0c\u76f4\u5230\u8f93\u5165\u7a7a\u95f4\uff0c\u5e76\u5c06\u5206\u89e3\u5206\u6570\u4f5c\u4e3a\u91cd\u8981\u6027\u5206\u6570\u3002<\/span><\/section>\n<\/li>\n
          • \n
            \u57fa\u4e8e\u4ee3\u7406\u7684\u65b9\u6cd5[56]\uff0c[57]\uff0c[58]\uff0c\u9996\u5148\u4ece\u7ed9\u5b9a\u4f8b\u5b50\u7684\u90bb\u5c45\u4e2d\u62bd\u53d6\u4e00\u4e2a\u6570\u636e\u96c6\u7684\u6837\u672c\u3002\u63a5\u4e0b\u6765\u5bf9\u91c7\u6837\u7684\u6570\u636e\u96c6\u5408\u62df\u5408\u4e00\u4e2a\u7b80\u5355\u4e14\u53ef\u89e3\u91ca\u7684\u6a21\u578b\uff0c\u5982\u51b3\u7b56\u6811\u3002\u901a\u8fc7\u89e3\u91ca\u4ee3\u7406\u6a21\u578b\u5b9e\u73b0\u5bf9\u539f\u59cb\u9884\u6d4b\u7684\u89e3\u91ca\u3002<\/span><\/section>\n<\/li>\n<\/ul>\n

            2\uff09\u6a21\u578b\u7ea7\u65b9\u6cd5<\/span><\/h4>\n
            \u6a21\u578b\u7ea7\u65b9\u6cd5\u76f4\u63a5\u89e3\u91ca\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u6a21\u578b\uff0c\u4e0d\u8003\u8651\u4efb\u4f55\u5177\u4f53\u7684\u8f93\u5165\u5b9e\u4f8b\u3002\u8fd9\u79cd input-independent \u7684\u89e3\u91ca\u662f\u9ad8\u5c42\u6b21\u7684\uff0c\u80fd\u591f\u89e3\u91ca\u4e00\u822c\u6027\u884c\u4e3a\u3002\u4e0e\u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u76f8\u6bd4\uff0c\u8fd9\u4e2a\u65b9\u5411\u7684\u63a2\u7d22\u8fd8\u6bd4\u8f83\u5c11\u3002\u73b0\u6709\u7684\u6a21\u578b\u7ea7\u65b9\u6cd5\u53ea\u6709XGNN[41]\uff0c\u5b83\u662f\u57fa\u4e8e\u56fe\u751f\u6210\u7684\uff0c\u901a\u8fc7\u751f\u6210 graph patterns\u4f7f\u67d0\u4e00\u7c7b\u7684\u9884\u6d4b\u6982\u7387\u6700\u5927\u5316\uff0c\u5e76\u5229\u7528 graph patterns \u6765\u89e3\u91ca\u8fd9\u4e00\u7c7b\u3002<\/span><\/section>\n
            \u603b\u7684\u6765\u8bf4\uff0c\u8fd9\u4e24\u7c7b\u65b9\u6cd5\u4ece\u4e0d\u540c\u7684\u89d2\u5ea6\u89e3\u91ca\u4e86\u6df1\u5ea6\u56fe\u6a21\u578b\u3002<\/span>\u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u63d0\u4f9b\u4e86\u9488\u5bf9\u5177\u4f53\u5b9e\u4f8b\u7684\u89e3\u91ca\uff0c\u800c\u6a21\u578b\u7ea7\u65b9\u6cd5\u5219\u63d0\u4f9b\u4e86\u9ad8\u5c42\u6b21\u7684\u89c1\u89e3\u548c\u5bf9\u6df1\u5ea6\u56fe\u6a21\u578b\u5de5\u4f5c\u539f\u7406\u7684\u4e00\u822c\u7406\u89e3\u3002<\/span><\/strong><\/section>\n
            \n

            <\/figcaption><\/figure>\n<\/section>\n
            “Type “\u8868\u793a\u63d0\u4f9b\u4ec0\u4e48\u7c7b\u578b\u7684\u89e3\u91ca\uff0c”Learning “\u8868\u793a\u662f\u5426\u6d89\u53ca\u5b66\u4e60\u8fc7\u7a0b\uff0c”Task “\u8868\u793a\u6bcf\u79cd\u65b9\u6cd5\u53ef\u4ee5\u5e94\u7528\u4e8e\u4ec0\u4e48\u4efb\u52a1<\/span>\uff08GC\u8868\u793a\u56fe\u5206\u7c7b\uff0cNC\u8868\u793a\u8282\u70b9\u5206\u7c7b\uff09<\/span>\uff0c”Target “\u8868\u793a\u89e3\u91ca\u7684\u5bf9\u8c61<\/span>\uff08N\u8868\u793a\u8282\u70b9\uff0cE\u8868\u793a\u8fb9\uff0cNF\u8868\u793a\u8282\u70b9\u7279\u5f81\uff0cWalk\u8868\u793a\u56fe\u6e38\u8d70\uff09<\/span>\uff0c”Black-box “\u8868\u793a\u5728\u89e3\u91ca\u9636\u6bb5\u662f\u5426\u5c06\u8bad\u7ec3\u597d\u7684GNNs\u4f5c\u4e3a\u9ed1\u76d2\u5904\u7406\uff0c”Flow “\u8868\u793a\u89e3\u91ca\u7684\u8ba1\u7b97\u6d41\u7a0b\uff0c”Design “\u8868\u793a\u89e3\u91ca\u65b9\u6cd5\u662f\u5426\u6709\u9488\u5bf9\u56fe\u6570\u636e\u7684\u5177\u4f53\u8bbe\u8ba1\u3002<\/span><\/section>\n
            <\/p>\n
            \n

             <\/span><\/p>\n<\/section>\n

            <\/span><\/section>\n

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

            <\/section>\n
            \n
            <\/svg><\/section>\n<\/section>\n<\/section>\n
            \n

            3. \u65b9\u6cd5\u4ecb\u7ecd<\/strong><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

            <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

            3.1 \u57fa\u4e8e\u68af\u5ea6\/\u7279\u5f81\u7684\u65b9\u6cd5\uff08Gradients\/Features-Based Methods\uff09<\/span><\/h3>\n
            \u91c7\u7528\u68af\u5ea6\u6216\u7279\u5f81\u6765\u89e3\u91ca\u6df1\u5ea6\u6a21\u578b\u662f\u6700\u76f4\u63a5\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u5728\u56fe\u50cf\u548c\u6587\u672c\u4efb\u52a1\u4e2d\u88ab\u5e7f\u6cdb\u4f7f\u7528\u3002<\/span>\u5176\u5173\u952e\u601d\u60f3\u662f\u5c06\u68af\u5ea6\u6216\u9690\u85cf\u7684\u7279\u5f81\u56fe\u503c\u4f5c\u4e3a\u8f93\u5165\u91cd\u8981\u6027\u7684\u8fd1\u4f3c\u503c<\/span><\/strong>\u3002<\/strong>\u4e00\u822c\u6765\u8bf4\uff0c\u5728\u8fd9\u7c7b\u65b9\u6cd5\u4e2d\uff0c\u68af\u5ea6\u6216\u7279\u5f81\u503c\u8d8a\u5927\uff0c\u8868\u793a\u91cd\u8981\u6027\u8d8a\u9ad8\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c<\/span>\u68af\u5ea6\u548c\u9690\u85cf\u7279\u5f81\u90fd\u4e0e\u6a21\u578b\u53c2\u6570\u9ad8\u5ea6\u76f8\u5173\uff0c\u90a3\u4e48\u8fd9\u6837\u7684\u89e3\u91ca\u53ef\u4ee5\u53cd\u6620\u51fa\u6a21\u578b\u6240\u5305\u542b\u7684\u4fe1\u606f\u3002<\/span><\/strong>\u672c\u6587\u5c06\u4ecb\u7ecd\u6700\u8fd1\u63d0\u51fa\u7684\u51e0\u79cd\u65b9\u6cd5\uff0c\u5305\u62ec\uff1aSA[49]\u3001Guided BP[49]\u3001CAM[50]\u548cGrad-CAM[50]\u3002<\/span>\u8fd9\u4e9b\u65b9\u6cd5\u7684\u5173\u952e\u533a\u522b\u5728\u4e8e\u68af\u5ea6\u53cd\u5411\u4f20\u64ad\u7684\u8fc7\u7a0b\u4ee5\u53ca\u5982\u4f55\u5c06\u4e0d\u540c\u7684\u9690\u85cf\u7279\u5f81\u56fe\u7ed3\u5408\u8d77\u6765\u3002<\/span><\/strong><\/section>\n

            1\uff09SA<\/span><\/strong><\/h4>\n
            SA[49]\u76f4\u63a5\u91c7\u7528\u68af\u5ea6\u7684\u5e73\u65b9\u503c\u4f5c\u4e3a\u4e0d\u540c\u8f93\u5165\u7279\u5f81\u7684\u91cd\u8981\u6027\u5f97\u5206\u3002\u8f93\u5165\u7279\u5f81\u53ef\u4ee5\u662f\u56fe\u8282\u70b9\u3001\u8fb9\u6216\u8282\u70b9\u7279\u5f81\u3002\u5b83\u5047\u8bbe\u7edd\u5bf9\u68af\u5ea6\u503c\u8d8a\u9ad8\uff0c\u8bf4\u660e\u76f8\u5e94\u7684\u8f93\u5165\u7279\u5f81\u8d8a\u91cd\u8981\u3002\u867d\u7136\u5b83\u7b80\u5355\u9ad8\u6548\uff0c\u4f46\u6709\u51e0\u4e2a\u5c40\u9650\u6027\uff1a1\uff09SA\u65b9\u6cd5\u53ea\u80fd\u53cd\u6620\u8f93\u5165\u548c\u8f93\u51fa\u4e4b\u95f4\u7684\u654f\u611f\u7a0b\u5ea6\uff0c\u4e0d\u80fd\u5f88\u51c6\u786e\u5730\u8868\u8fbe\u91cd\u8981\u6027<\/span>\uff08\u654f\u611f\u5ea6\u4e0d\u7b49\u4e8e\u91cd\u8981\u6027\uff09<\/span>\u30022\uff09\u8fd8\u5b58\u5728\u9971\u548c\u95ee\u9898[59]\u3002\u5373\u5728\u6a21\u578b\u6027\u80fd\u8fbe\u5230\u9971\u548c\u7684\u533a\u57df\uff0c\u5176\u8f93\u51fa\u76f8\u5bf9\u4e8e\u4efb\u4f55\u8f93\u5165\u53d8\u5316\u7684\u53d8\u5316\u90fd\u662f\u5341\u5206\u5fae\u5c0f\u7684\uff0c\u68af\u5ea6\u5f88\u96be\u53cd\u6620\u8f93\u5165\u7684\u8d21\u732e\u7a0b\u5ea6\u3002<\/span><\/section>\n

            2\uff09Guided BP<\/span><\/h4>\n
            Guided BP[49]\u4e0eSA\u6709\u7740\u76f8\u4f3c\u7684\u601d\u60f3\uff0c\u4f46\u4fee\u6539\u4e86\u53cd\u5411\u4f20\u64ad\u68af\u5ea6\u7684\u8fc7\u7a0b\u3002\u7531\u4e8e\u8d1f\u68af\u5ea6\u5f88\u96be\u89e3\u91ca\uff0cGuided BP\u53ea\u53cd\u5411\u4f20\u64ad\u6b63\u68af\u5ea6\uff0c\u800c\u5c06\u8d1f\u68af\u5ea6\u526a\u6210\u96f6\u3002\u56e0\u6b64Guided BP\u4e0eSA\u6709\u7740\u76f8\u540c\u7684\u5c40\u9650\u6027\u3002<\/span><\/section>\n

            3\uff09CAM<\/span><\/h4>\n
            CAM [50] \u5c06\u6700\u540e\u4e00\u5c42\u7684\u8282\u70b9\u7279\u5f81\u6620\u5c04\u5230\u8f93\u5165\u7a7a\u95f4\uff0c\u4ece\u800c\u8bc6\u522b\u91cd\u8981\u8282\u70b9\u3002\u5b83\u8981\u6c42GNN\u6a21\u578b\u91c7\u7528\u5168\u5c40\u5e73\u5747\u6c60\u5316\u5c42\u548c\u5168\u8fde\u63a5\u5c42\u4f5c\u4e3a\u6700\u7ec8\u5206\u7c7b\u5668\u3002CAM\u5c06\u6700\u7ec8\u7684\u8282\u70b9\u5d4c\u5165\uff0c\u901a\u8fc7\u52a0\u6743\u6c42\u548c\u7684\u65b9\u5f0f\u7ec4\u5408\u4e0d\u540c\u7684\u7279\u5f81\u56fe\uff0c\u4ece\u800c\u83b7\u5f97\u8f93\u5165\u8282\u70b9\u7684\u91cd\u8981\u6027\u5206\u6570\u3002\u6743\u91cd\u662f\u4ece\u4e0e\u76ee\u6807\u9884\u6d4b\u8fde\u63a5\u7684\u6700\u7ec8\u5168\u8fde\u63a5\u5c42\u83b7\u5f97\u7684\u3002\u8be5\u65b9\u6cd5\u975e\u5e38\u7b80\u5355\u9ad8\u6548\uff0c\u4f46\u4ecd\u6709\u51e0\u5927\u9650\u5236\uff1a1\uff09CAM\u5bf9GNN\u7ed3\u6784\u6709\u7279\u6b8a\u8981\u6c42\uff0c\u9650\u5236\u4e86\u5b83\u7684\u5e94\u7528\u548c\u63a8\u5e7f\u30022\uff09\u5b83\u5047\u8bbe\u6700\u7ec8\u7684\u8282\u70b9\u5d4c\u5165\u53ef\u4ee5\u53cd\u6620\u8f93\u5165\u7684\u91cd\u8981\u6027\uff0c\u8fd9\u662f\u542f\u53d1\u5f0f\u7684\uff0c\u53ef\u80fd\u4e0d\u662f\u771f\u7684\u30023\uff09\u5b83\u53ea\u80fd\u89e3\u91ca\u56fe\u5206\u7c7b\u6a21\u578b\uff0c\u4e0d\u80fd\u5e94\u7528\u4e8e\u8282\u70b9\u5206\u7c7b\u4efb\u52a1\u4e2d\u3002<\/span><\/section>\n

            4\uff09Grad-CAM<\/span><\/h4>\n
            Grad-CAM [50] \u901a\u8fc7\u53bb\u9664\u5168\u5c40\u5e73\u5747\u6c60\u5316\u5c42\u7684\u7ea6\u675f\uff0c\u5c06CAM\u6269\u5c55\u5230\u4e00\u822c\u56fe\u5206\u7c7b\u6a21\u578b\u3002\u540c\u6837\uff0c\u5b83\u4e5f\u5c06\u6700\u7ec8\u7684\u8282\u70b9\u5d4c\u5165\u6620\u5c04\u5230\u8f93\u5165\u7a7a\u95f4\u6765\u8861\u91cf\u8282\u70b9\u91cd\u8981\u6027\u3002\u4f46\u662f\uff0c\u5b83\u6ca1\u6709\u4f7f\u7528\u5168\u5c40\u5e73\u5747\u6c60\u5316\u8f93\u51fa\u548c\u5168\u8fde\u63a5\u5c42\u8f93\u51fa\u4e4b\u95f4\u7684\u6743\u91cd\uff0c\u800c\u662f\u91c7\u7528\u68af\u5ea6\u4f5c\u4e3a\u6743\u91cd\u6765\u7ec4\u5408\u4e0d\u540c\u7684\u7279\u5f81\u56fe\u3002\u4e0eCAM\u76f8\u6bd4\uff0cGrad-CAM\u4e0d\u9700\u8981GNN\u6a21\u578b\u5728\u6700\u7ec8\u7684\u5168\u8fde\u63a5\u5c42\u4e4b\u524d\u91c7\u7528\u5168\u5c40\u5e73\u5747\u6c60\u5316\u5c42\u3002\u4f46\u5b83\u4e5f\u662f\u57fa\u4e8e\u542f\u53d1\u5f0f\u5047\u8bbe\uff0c\u65e0\u6cd5\u89e3\u91ca\u8282\u70b9\u5206\u7c7b\u6a21\u578b\u3002<\/span><\/section>\n

            3.2 \u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\uff08Perturbation-Based Methods\uff09<\/span><\/h3>\n
            \u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5[14]\uff0c[15]\uff0c[60]\u88ab\u5e7f\u6cdb\u7528\u4e8e\u89e3\u91ca\u6df1\u5ea6\u56fe\u50cf\u6a21\u578b\u3002<\/span>\u5176\u6839\u672c\u52a8\u673a\u662f\u7814\u7a76\u4e0d\u540c\u8f93\u5165\u6270\u52a8\u4e0b\u7684\u8f93\u51fa\u53d8\u5316\u3002<\/span><\/strong>\u5f53\u91cd\u8981\u7684\u8f93\u5165\u4fe1\u606f\u88ab\u4fdd\u7559<\/span>\uff08\u6ca1\u6709\u88ab\u6270\u52a8\uff09<\/span>\u65f6\uff0c\u9884\u6d4b\u7ed3\u679c\u5e94\u8be5\u4e0e\u539f\u59cb\u9884\u6d4b\u7ed3\u679c\u76f8\u4f3c\u3002\u8bba\u6587 [14]\uff0c[15]\uff0c[60]\u5b66\u4e60\u4e00\u4e2a\u751f\u6210\u5668\u6765\u751f\u6210\u63a9\u7801\uff0c\u4ee5\u9009\u62e9\u91cd\u8981\u7684\u8f93\u5165\u50cf\u7d20\u6765\u89e3\u91ca\u6df1\u5ea6\u56fe\u50cf\u6a21\u578b\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u80fd\u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u6a21\u578b\uff0c\u56fe\u6570\u636e\u662f\u4ee5\u8282\u70b9\u548c\u8fb9\u6765\u8868\u793a\u7684\uff0c\u5b83\u4eec\u4e0d\u80fd\u8c03\u6574\u5927\u5c0f\u4ee5\u5171\u4eab\u76f8\u540c\u7684\u8282\u70b9\u548c\u8fb9\u6570\uff0c\u7ed3\u6784\u4fe1\u606f\u5bf9\u56fe\u6765\u8bf4\u81f3\u5173\u91cd\u8981\uff0c\u53ef\u4ee5\u51b3\u5b9a\u56fe\u7684\u529f\u80fd\u3002<\/span><\/section>\n
            \n
            <\/figure>\n<\/section>\n
            \u56fe2 \u57fa\u4e8e\u6270\u52a8\u65b9\u6cd5\u7684\u4e00\u822c\u6d41\u7a0b<\/span><\/section>\n
            \u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\u91c7\u7528\u4e0d\u540c\u7684\u63a9\u7801\u751f\u6210\u7b97\u6cd5\u6765\u83b7\u5f97\u4e0d\u540c\u7c7b\u578b\u7684\u63a9\u7801\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c<\/span>\u63a9\u7801\u53ef\u4ee5\u5bf9\u5e94\u8282\u70b9\u3001\u8fb9\u6216\u8282\u70b9\u7279\u5f81\u3002<\/span><\/strong>\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u4e00\u4e2a\u8282\u70b9\u7279\u5f81\u7684\u8f6f\u63a9\u7801<\/strong>\uff0c\u4e00\u4e2a\u8fb9\u7684\u79bb\u6563\u63a9\u7801<\/strong>\u548c\u4e00\u4e2a\u8282\u70b9\u7684\u8fd1\u4f3c\u79bb\u6563\u63a9\u7801<\/strong>\u3002\u7136\u540e\uff0c\u5c06\u63a9\u7801\u4e0e\u8f93\u5165\u56fe\u7ed3\u5408\u8d77\u6765\uff0c\u5f97\u5230\u4e00\u4e2a\u5305\u542b\u91cd\u8981\u8f93\u5165\u4fe1\u606f\u7684\u65b0\u56fe<\/strong>\uff0c\u906e\u853d\u6389\u4e0d\u9700\u8981\u7684\u4fe1\u606f\u3002\u6700\u7ec8\u5c06\u65b0\u56fe\u8f93\u5165\u5230\u8bad\u7ec3\u597d\u7684GNN\u4e2d\uff0c\u8bc4\u4f30\u63a9\u7801\u5e76\u66f4\u65b0\u63a9\u7801\u751f\u6210\u7b97\u6cd5\u3002<\/strong>\u672c\u6587\u5c06\u4ecb\u7ecd\u51e0\u79cd\u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\uff0c\u5305\u62ec\uff1aGNNExplainer[42]\u3001PGExplainer[43]\u3001ZORRO[51]\u3001GraphMask[52]\uff0cCausal Screening[53]\u3002\u76f4\u89c2\u5730\u8bb2\uff0c\u63a9\u7801\u6355\u6349\u5230\u7684\u91cd\u8981\u8f93\u5165\u7279\u5f81\u5e94\u8be5\u4f20\u8fbe\u5173\u952e\u7684\u8bed\u4e49\u610f\u4e49\uff0c\u4ece\u800c\u5f97\u5230\u4e0e\u539f\u56fe\u76f8\u4f3c\u7684\u9884\u6d4b\u7ed3\u679c\u3002<\/strong>\u8fd9\u4e9b\u65b9\u6cd5\u7684\u533a\u522b\u4e3b\u8981\u5728\u4e8e\u4e09\u4e2a\u65b9\u9762\uff1a\u63a9\u7801\u751f\u6210\u7b97\u6cd5\u3001\u63a9\u7801\u7c7b\u578b\u548c\u76ee\u6807\u51fd\u6570\u3002<\/strong><\/span><\/section>\n
            \u8f6f\u63a9\u7801\u5305\u542b[0\uff0c1]\u4e4b\u95f4\u7684\u8fde\u7eed\u503c\uff0c\u63a9\u7801\u751f\u6210\u7b97\u6cd5\u53ef\u4ee5\u76f4\u63a5\u901a\u8fc7\u53cd\u5411\u4f20\u64ad\u8fdb\u884c\u66f4\u65b0\u3002\u4f46\u8f6f\u63a9\u7801\u5b58\u5728 “introduced evidence “\u7684\u95ee\u9898[14]\uff0c\u5373\u63a9\u7801\u4e2d\u4efb\u4f55\u975e\u96f6\u6216\u975e\u4e00\u7684\u503c\u90fd\u53ef\u80fd\u7ed9\u8f93\u5165\u56fe\u5f15\u5165\u65b0\u7684\u8bed\u4e49\u6216\u65b0\u7684\u566a\u58f0<\/strong>\uff0c\u4ece\u800c\u5f71\u54cd\u89e3\u91ca\u7ed3\u679c\u3002\u540c\u65f6\uff0c\u79bb\u6563\u63a9\u7801\u53ea\u5305\u542b\u79bb\u6563\u503c0\u548c1\uff0c\u7531\u4e8e\u6ca1\u6709\u5f15\u5165\u65b0\u7684\u6570\u503c\uff0c\u53ef\u4ee5\u907f\u514d  “introduced evidence “\u95ee\u9898\u3002\u4f46\u662f\uff0c\u79bb\u6563\u63a9\u7801\u603b\u662f\u6d89\u53ca\u5230\u4e0d\u53ef\u5fae\u7684\u95ee\u9898<\/strong>\uff0c\u5982\u91c7\u6837\u3002\u4e3b\u6d41\u7684\u89e3\u51b3\u65b9\u6cd5\u662f\u7b56\u7565\u68af\u5ea6\u6280\u672f[61]\u3002\u8bba\u6587[45]\uff0c[62]\uff0c[63]\u63d0\u51fa\u91c7\u7528\u91cd\u53c2\u6570\u5316\u6280\u5de7\uff0c\u5982Gumbel-Softmax\u4f30\u8ba1\u548c\u7a00\u758f\u677e\u5f1b\uff0c\u6765\u903c\u8fd1\u79bb\u6563\u63a9\u7801\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8f93\u51fa\u7684\u63a9\u7801\u5e76\u4e0d\u662f\u4e25\u683c\u610f\u4e49\u4e0a\u7684\u79bb\u6563\u63a9\u7801\uff0c\u800c\u662f\u63d0\u4f9b\u4e86\u4e00\u4e2a\u5f88\u597d\u7684\u8fd1\u4f3c\u503c\uff0c\u8fd9\u4e0d\u4ec5\u53ef\u4ee5\u5b9e\u73b0\u53cd\u5411\u4f20\u64ad\uff0c\u800c\u4e14\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u7f13\u89e3\u4e86\u201dintroduced evidence\u201c\u7684\u95ee\u9898\u3002<\/span><\/section>\n
            \u63a5\u4e0b\u6765\u5c06\u8be6\u7ec6\u7684\u4ecb\u7ecd\u76ee\u524d\u5b58\u5728\u7684\u51e0\u79cd\u57fa\u4e8e\u6270\u52a8\u7684\u89e3\u91ca\u6027\u65b9\u6cd5\uff0c\u6211\u4eec\u53ef\u4ee5\u91cd\u70b9\u5173\u6ce8\u5b83\u4eec\u7684\u4f5c\u7528\u5bf9\u8c61<\/span>\uff08\u8282\u70b9\u6216\u8fb9\u6216\u8282\u70b9\u7279\u5f81\uff09<\/span>\uff0c\u4ee5\u53ca\u5b83\u4eec\u7684\u63a9\u7801\u751f\u6210\u7b97\u6cd5\u3001\u63a9\u7801\u7c7b\u578b\u548c\u76ee\u6807\u51fd\u6570\u3002<\/span><\/section>\n

            1\uff09GNNExplainer<\/span><\/h4>\n
            GNNExplainer [42] \u5b66\u4e60<\/span>\u8fb9\u548c\u8282\u70b9\u7279\u5f81<\/span><\/strong>\u7684\u8f6f\u63a9\u7801\uff0c\u901a\u8fc7\u63a9\u7801\u4f18\u5316\u6765\u89e3\u91ca\u9884\u6d4b\u3002\u8f6f\u63a9\u7801\u88ab\u968f\u673a\u521d\u59cb\u5316\uff0c\u5e76\u88ab\u89c6\u4e3a\u53ef\u8bad\u7ec3\u53d8\u91cf\u3002\u7136\u540e\u901a\u8fc7\u5143\u7d20\u70b9\u4e58\u5c06\u63a9\u7801\u4e0e\u539f\u59cb\u56fe\u7ed3\u5408\u3002\u6700\u5927\u5316\u539f\u59cb\u56fe\u7684\u9884\u6d4b\u548c\u65b0\u83b7\u5f97\u7684\u56fe\u7684\u9884\u6d4b\u4e4b\u95f4\u7684\u4e92\u4fe1\u606f\u6765\u4f18\u5316\u63a9\u7801\u3002<\/strong>\u4f46\u5f97\u5230\u7684\u63a9\u7801\u4ecd\u7136\u662f\u8f6f\u63a9\u7801\uff0c\u56e0\u6b64\u65e0\u6cd5\u907f\u514d “introduced evidence”\u95ee\u9898\u3002\u6b64\u5916\uff0c\u63a9\u7801\u662f\u9488\u5bf9\u6bcf\u4e2a\u8f93\u5165\u56fe\u5355\u72ec\u4f18\u5316\u7684\uff0c\u56e0\u6b64\u89e3\u91ca\u53ef\u80fd\u7f3a\u4e4f\u5168\u5c40\u89c6\u89d2\u3002<\/span><\/section>\n

            2\uff09PGExplainer<\/span><\/h4>\n
            PGExplainer[43]\u5b66\u4e60\u8fb9<\/strong>\u7684\u8fd1\u4f3c\u79bb\u6563\u63a9\u7801\u6765\u89e3\u91ca\u9884\u6d4b\u3002\u5b83\u8bad\u7ec3\u4e00\u4e2a\u53c2\u6570\u5316\u7684\u63a9\u7801\u9884\u6d4b\u5668\u6765\u9884\u6d4b\u8fb9\u63a9\u7801\u3002<\/strong>\u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u56fe\uff0c\u9996\u5148\u901a\u8fc7\u62fc\u63a5\u8282\u70b9\u5d4c\u5165\u6765\u83b7\u5f97\u6bcf\u4e2a\u8fb9\u7684\u5d4c\u5165\uff0c\u7136\u540e\u9884\u6d4b\u5668\u4f7f\u7528\u8fb9\u5d4c\u5165\u6765\u9884\u6d4b\u8fb9\u63a9\u7801\u3002\u9884\u6d4b\u5668\u4f7f\u7528\u8fb9\u5d4c\u5165\u6765\u9884\u6d4b\u6bcf\u4e2a\u8fb9\u88ab\u9009\u4e2d\u7684\u6982\u7387<\/span>\uff08\u88ab\u89c6\u4e3a\u91cd\u8981\u6027\u5206\u6570\uff09<\/span>\u3002\u901a\u8fc7\u91cd\u53c2\u6570\u5316\u6280\u5de7\u5bf9\u8fd1\u4f3c\u7684\u79bb\u6563\u63a9\u7801\u8fdb\u884c\u91c7\u6837\u3002\u6700\u540e\u901a\u8fc7\u6700\u5927\u5316\u539f\u59cb\u9884\u6d4b\u548c\u65b0\u9884\u6d4b\u4e4b\u95f4\u7684\u76f8\u4e92\u4fe1\u606f\u6765\u8bad\u7ec3\u63a9\u7801\u9884\u6d4b\u5668\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5373\u4f7f\u91c7\u7528\u4e86\u91cd\u53c2\u6570\u5316\u6280\u5de7\uff0c\u5f97\u5230\u7684\u63a9\u7801\u5e76\u4e0d\u662f\u4e25\u683c\u610f\u4e49\u4e0a\u7684\u79bb\u6563\u63a9\u7801\uff0c\u4f46\u53ef\u4ee5\u5f88\u5927\u7a0b\u5ea6\u4e0a\u7f13\u89e3 “introduced evidence”\u7684\u95ee\u9898\u3002\u7531\u4e8e\u6570\u636e\u96c6\u4e2d\u7684\u6240\u6709\u8fb9\u90fd\u5171\u4eab\u76f8\u540c\u7684\u9884\u6d4b\u5668\uff0c\u56e0\u6b64\u89e3\u91ca\u53ef\u4ee5\u63d0\u4f9b\u5bf9\u8bad\u7ec3\u597d\u7684GNN\u7684\u5168\u5c40\u7406\u89e3\u3002<\/span><\/section>\n

            3\uff09GraphMask<\/span><\/h4>\n
            GraphMask[52]\u662f\u4e00\u79cd\u4e8b\u540e\u89e3\u91caGNN\u5404\u5c42\u4e2d\u8fb9<\/strong>\u91cd\u8981\u6027\u7684\u65b9\u6cd5\u3002\u4e0ePGExplainer\u7c7b\u4f3c\uff0c\u5b83\u8bad\u7ec3\u4e00\u4e2a\u5206\u7c7b\u5668\u6765\u9884\u6d4b\u662f\u5426\u53ef\u4ee5\u4e22\u5f03\u4e00\u6761\u8fb9\u800c\u4e0d\u5f71\u54cd\u539f\u6765\u7684\u9884\u6d4b<\/strong>\u3002\u7136\u800c\uff0cGraphMask\u4e3a\u6bcf\u4e00\u5c42GNN\u83b7\u53d6\u8fb9\u63a9\u7801\uff0c\u800cPGExplainer\u53ea\u5173\u6ce8\u8f93\u5165\u7a7a\u95f4\u3002\u6b64\u5916\uff0c\u4e3a\u4e86\u907f\u514d\u6539\u53d8\u56fe\u7ed3\u6784\uff0c\u88ab\u4e22\u5f03\u7684\u8fb9\u88ab\u53ef\u5b66\u4e60\u7684\u57fa\u7ebf\u8fde\u63a5\u6240\u53d6\u4ee3\uff0c\u57fa\u7ebf\u8fde\u63a5\u662f\u4e0e\u8282\u70b9\u5d4c\u5165\u76f8\u540c\u7ef4\u5ea6\u7684\u5411\u91cf\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u91c7\u7528\u4e8c\u8fdb\u5236Concrete\u5206\u5e03[63]\u548c\u91cd\u53c2\u6570\u5316\u6280\u5de7\u6765\u8fd1\u4f3c\u79bb\u6563\u63a9\u7801\u3002\u6b64\u5916\uff0c\u5206\u7c7b\u5668\u4f7f\u7528\u6574\u4e2a\u6570\u636e\u96c6\u901a\u8fc7\u6700\u5c0f\u5316\u4e00\u4e2a\u6563\u5ea6\u9879\u6765\u8bad\u7ec3\uff0c\u7528\u4e8e\u8861\u91cf\u7f51\u7edc\u9884\u6d4b\u4e4b\u95f4\u7684\u5dee\u5f02\u3002\u4e0ePGExplainer\u7c7b\u4f3c\uff0c\u5b83\u53ef\u4ee5\u5f88\u5927\u7a0b\u5ea6\u4e0a\u7f13\u89e3 “introduced evidence”\u95ee\u9898\uff0c\u5e76\u5bf9\u8bad\u7ec3\u540e\u7684GNN\u8fdb\u884c\u5168\u5c40\u7406\u89e3\u3002<\/span><\/section>\n

            3\uff09ZORRO<\/span><\/h4>\n
            ZORRO[51]\u91c7\u7528\u79bb\u6563\u63a9\u7801\u6765\u8bc6\u522b\u91cd\u8981\u7684\u8f93\u5165\u8282\u70b9\u548c\u8282\u70b9\u7279\u5f81<\/strong>\u3002\u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u56fe\uff0c\u91c7\u7528\u8d2a\u5fc3\u7b97\u6cd5\u9010\u6b65\u9009\u62e9\u8282\u70b9\u6216\u8282\u70b9\u7279\u5f81\u3002\u6bcf\u4e00\u6b65\u90fd\u4f1a\u9009\u62e9\u4e00\u4e2afidelity score\u6700\u9ad8\u7684\u8282\u70b9\u6216\u4e00\u4e2a\u8282\u70b9\u7279\u5f81\u3002\u901a\u8fc7\u56fa\u5b9a\u6240\u9009\u8282\u70b9\/\u7279\u5f81\uff0c\u5e76\u7528\u968f\u673a\u566a\u58f0\u503c\u66ff\u6362\u5176\u4ed6\u8282\u70b9\/\u7279\u5f81\uff0c\u6765\u8861\u91cf\u65b0\u7684\u9884\u6d4b\u4e0e\u6a21\u578b\u539f\u59cb\u9884\u6d4b\u7684\u5339\u914d\u7a0b\u5ea6\u3002\u7531\u4e8e\u4e0d\u6d89\u53ca\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u56e0\u6b64\u907f\u514d\u4e86\u79bb\u6563\u63a9\u7801\u7684\u4e0d\u53ef\u5fae\u9650\u5236\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u4f7f\u7528\u786c\u63a9\u7801\uff0cZORRO\u4e0d\u4f1a\u53d7\u5230 “introduced evidence “\u95ee\u9898\u7684\u5f71\u54cd\u3002\u7136\u800c\uff0c\u8d2a\u5a6a\u7684\u63a9\u7801\u9009\u62e9\u7b97\u6cd5\u53ef\u80fd\u5bfc\u81f4\u5c40\u90e8\u6700\u4f18\u89e3\u91ca\u3002\u6b64\u5916\uff0c\u7531\u4e8e\u63a9\u7801\u662f\u4e3a\u6bcf\u4e2a\u56fe\u5f62\u5355\u72ec\u751f\u6210\u7684\uff0c\u56e0\u6b64\u89e3\u91ca\u53ef\u80fd\u7f3a\u4e4f\u5168\u5c40\u7684\u7406\u89e3\u3002<\/span><\/section>\n

            4\uff09Causal Screening<\/span><\/h4>\n
            Causal Screening[53]\u7814\u7a76\u8f93\u5165\u56fe\u4e2d\u4e0d\u540c\u8fb9<\/strong>\u7684\u56e0\u679c\u5f52\u56e0\u3002\u5b83\u4e3a explanatory subgraph \u786e\u5b9a\u4e00\u4e2a\u8fb9\u63a9\u7801\u3002\u5173\u952e\u601d\u60f3\u662f\uff1a\u7814\u7a76\u5728\u5f53\u524d explanatory subgraph \u4e2d\u589e\u52a0\u4e00\u6761\u8fb9\u65f6\u9884\u6d4b\u7684\u53d8\u5316\uff0c\u5373\u6240\u8c13\u7684\u56e0\u679c\u6548\u5e94\u3002<\/strong>\u5bf9\u4e8e\u6bcf\u4e00\u6b65\uff0c\u5b83\u90fd\u4f1a\u7814\u7a76\u4e0d\u540c\u8fb9\u7684\u56e0\u679c\u6548\u5e94\uff0c\u5e76\u9009\u62e9\u4e00\u6761\u8fb9\u6dfb\u52a0\u5230\u5b50\u56fe\u4e2d\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u91c7\u7528\u4e2a\u4f53\u56e0\u679c\u6548\u5e94(ICE)\u6765\u9009\u62e9\u8fb9\uff0c\u5373\u6d4b\u91cf\u5728\u5b50\u56fe\u4e2d\u6dfb\u52a0\u4e0d\u540c\u8fb9\u540e\u7684\u4e92\u4fe1\u606f(\u539f\u56fe\u4e0e\u89e3\u91ca\u5b50\u56fe\u7684\u9884\u6d4b\u4e4b\u95f4)\u5dee\u5f02\u3002\u4e0eZORRO\u7c7b\u4f3c\uff0cCausal Screening\u662f\u4e00\u79cd\u8d2a\u5fc3\u7b97\u6cd5\uff0c\u4e0d\u9700\u8981\u4efb\u4f55\u8bad\u7ec3\u8fc7\u7a0b\u5c31\u80fd\u751f\u6210\u79bb\u6563\u63a9\u7801\u3002\u56e0\u6b64\uff0c\u5b83\u4e0d\u4f1a\u53d7\u5230 “introduced evidence “\u95ee\u9898\u7684\u56f0\u6270\uff0c\u4f46\u53ef\u80fd\u7f3a\u4e4f\u5168\u5c40\u6027\u7684\u7406\u89e3\uff0c\u800c\u505c\u7559\u5728\u5c40\u90e8\u6700\u4f18\u89e3\u91ca\u4e0a\u3002<\/span><\/section>\n

            3.3 \u57fa\u4e8e\u4ee3\u7406\u7684\u65b9\u6cd5\uff08Surrogate Methods\uff09<\/span><\/h3>\n
            \u7531\u4e8e\u8f93\u5165\u7a7a\u95f4\u548c\u8f93\u51fa\u9884\u6d4b\u4e4b\u95f4\u7684\u590d\u6742\u548c\u975e\u7ebf\u6027\u5173\u7cfb\uff0c\u6df1\u5ea6\u6a21\u578b\u7684\u89e3\u91ca\u5177\u6709\u6311\u6218\u6027\u3002\u4ee3\u7406\u65b9\u6cd5\u80fd\u591f\u4e3a\u56fe\u50cf\u6a21\u578b\u63d0\u4f9b\u5b9e\u4f8b\u7ea7\u89e3\u91ca\u3002\u5176\u57fa\u672c\u601d\u60f3\u662f\u5316\u7e41\u4e3a\u7b80\uff0c\u65e2\u7136\u65e0\u6cd5\u89e3\u91ca\u539f\u59cb\u6df1\u5ea6\u56fe\u6a21\u578b\uff0c\u90a3\u4e48\u91c7\u7528\u4e00\u4e2a\u7b80\u5355\u4e14\u53ef\u89e3\u91ca\u7684\u4ee3\u7406\u6a21\u578b\u6765\u8fd1\u4f3c\u590d\u6742\u7684\u6df1\u5c42\u6a21\u578b\uff0c\u5b9e\u73b0\u8f93\u5165\u5b9e\u4f8b\u7684\u90bb\u8fd1\u533a\u57df\u9884\u6d4b\u3002<\/strong>\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u662f\u5047\u8bbe\u8f93\u5165\u5b9e\u4f8b\u7684\u90bb\u8fd1\u533a\u57df\u7684\u5173\u7cfb\u4e0d\u90a3\u4e48\u590d\u6742\uff0c\u53ef\u4ee5\u88ab\u4e00\u4e2a\u8f83\u7b80\u5355\u7684\u4ee3\u7406\u6a21\u578b\u5f88\u597d\u5730\u6355\u83b7\u3002\u7136\u540e\u901a\u8fc7\u53ef\u89e3\u91ca\u7684\u4ee3\u7406\u6a21\u578b\u7684\u6765\u89e3\u91ca\u539f\u59cb\u9884\u6d4b\u3002\u5c06\u4ee3\u7406\u65b9\u6cd5\u5e94\u7528\u5230\u56fe\u57df\u662f\u4e00\u4e2a\u6311\u6218\uff0c\u56e0\u4e3a\u56fe\u6570\u636e\u662f\u79bb\u6563\u7684\uff0c\u5305\u542b\u62d3\u6251\u4fe1\u606f\u3002\u90a3\u4e48\u5982\u4f55\u5b9a\u4e49\u8f93\u5165\u56fe\u7684\u76f8\u90bb\u533a\u57df\uff0c\u4ee5\u53ca\u4ec0\u4e48\u6837\u7684\u53ef\u89e3\u91ca\u4ee3\u7406\u6a21\u578b\u662f\u5408\u9002\u7684\uff0c\u90fd\u662f\u4e0d\u6e05\u695a\u7684\u3002<\/span><\/section>\n
            \n
            <\/figure>\n<\/section>\n
            \u56fe3 \u4ee3\u7406\u6a21\u578b\u7684\u4e00\u822c\u6846\u67b6<\/span><\/section>\n
            \u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u56fe\u53ca\u5176\u9884\u6d4b\uff0c\u5b83\u4eec\u9996\u5148\u5bf9\u4e00\u4e2a\u5c40\u90e8\u6570\u636e\u96c6\u8fdb\u884c\u91c7\u6837\uff0c\u4ee5\u8868\u793a\u76ee\u6807\u6570\u636e\u5468\u56f4\u7684\u5173\u7cfb\u3002\u7136\u540e\u5e94\u7528\u4e0d\u540c\u7684\u4ee3\u7406\u65b9\u6cd5\u6765\u62df\u5408\u5c40\u90e8\u6570\u636e\u96c6\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u4ee3\u7406\u6a21\u578b\u4e00\u822c\u90fd\u662f\u7b80\u5355\u4e14\u53ef\u89e3\u91ca\u7684ML\u6a21\u578b\u3002\u6700\u540e\uff0c\u4ee3\u7406\u6a21\u578b\u7684\u89e3\u91ca\u53ef\u4ee5\u770b\u4f5c\u662f\u5bf9\u539f\u59cb\u9884\u6d4b\u7684\u89e3\u91ca\u3002\u672c\u6587\u5c06\u4ecb\u7ecd\u6700\u8fd1\u63d0\u51fa\u7684\u51e0\u79cd\u4ee3\u7406\u65b9\u6cd5\uff0c\u5305\u62ec\uff1aGraphLime[56]\u3001RelEx[57]\u548cPGM-Explainer[58]\u3002\u8fd9\u4e9b\u65b9\u6cd5\u7684\u4e00\u822c\u6d41\u7a0b\u5982\u56fe3\u6240\u793a\u3002\u4e3a\u4e86\u89e3\u91ca\u7ed9\u5b9a\u8f93\u5165\u56fe\u7684\u9884\u6d4b\uff0c\u5b83\u4eec\u9996\u5148\u83b7\u5f97\u4e00\u4e2a\u5305\u542b\u591a\u4e2a\u76f8\u90bb\u6570\u636e\u5bf9\u8c61\u53ca\u5176\u9884\u6d4b\u7684\u5c40\u90e8\u6570\u636e\u96c6\u3002\u7136\u540e\u62df\u5408\u4e00\u4e2a\u53ef\u89e3\u91ca\u6a21\u578b\u6765\u5b66\u4e60\u5c40\u90e8\u6570\u636e\u96c6\u3002\u6765\u81ea\u53ef\u89e3\u91ca\u6a21\u578b\u7684\u89e3\u91ca\u88ab\u89c6\u4e3a\u539f\u59cb\u6a21\u578b\u5bf9\u8f93\u5165\u56fe\u7684\u89e3\u91ca\u3002\u4e0d\u540c\u4ee3\u7406\u6a21\u578b\u5173\u952e\u7684\u533a\u522b\u5728\u4e8e\u4e24\u4e2a\u65b9\u9762\uff1a\u5982\u4f55\u83b7\u5f97\u5c40\u90e8\u6570\u636e\u96c6\u548c\u9009\u62e9\u4ec0\u4e48\u4ee3\u7406\u6a21\u578b\u3002<\/strong><\/span><\/section>\n

            1\uff09GraphLime<\/span><\/h4>\n
            GraphLime[56]\u5c06LIME[64]\u7b97\u6cd5\u6269\u5c55\u5230\u6df1\u5ea6\u56fe\u6a21\u578b\uff0c\u5e76\u7814\u7a76\u4e0d\u540c\u8282\u70b9\u7279\u5f81<\/strong>\u5bf9\u8282\u70b9\u5206\u7c7b\u4efb\u52a1\u7684\u91cd\u8981\u6027\u3002\u7ed9\u5b9a\u8f93\u5165\u56fe\u4e2d\u7684\u4e00\u4e2a\u76ee\u6807\u8282\u70b9\uff0c\u5c06\u5176N-hop \u90bb\u5c45\u8282\u70b9\u53ca\u5176\u9884\u6d4b\u503c\u89c6\u4e3a\u5c40\u90e8\u6570\u636e\u96c6\uff0c\u5176\u4e2dN\u7684\u5408\u7406\u8bbe\u7f6e\u662f\u8bad\u7ec3\u7684GNN\u7684\u5c42\u6570\u3002\u7136\u540e\u91c7\u7528\u975e\u7ebf\u6027\u4ee3\u7406\u6a21\u578bHSIC Lasso[65]\u6765\u62df\u5408\u5c40\u90e8\u6570\u636e\u96c6\u3002\u6839\u636eHSIC Lasso\u4e2d\u4e0d\u540c\u7279\u5f81\u7684\u6743\u91cd\uff0c\u53ef\u4ee5\u9009\u62e9\u91cd\u8981\u7684\u7279\u5f81\u6765\u89e3\u91caHSIC Lasso\u7684\u9884\u6d4b\u7ed3\u679c\u3002\u8fd9\u4e9b\u88ab\u9009\u53d6\u7684\u7279\u5f81\u88ab\u8ba4\u4e3a\u662f\u5bf9\u539f\u59cbGNN\u9884\u6d4b\u7684\u89e3\u91ca\u3002<\/strong>\u4f46\u662f\uff0cGraphLime\u53ea\u80fd\u63d0\u4f9b\u8282\u70b9\u7279\u5f81\u7684\u89e3\u91ca\uff0c\u5374\u5ffd\u7565\u4e86\u8282\u70b9\u548c\u8fb9\u7b49\u56fe\u7ed3\u6784\uff0c\u800c\u8fd9\u4e9b\u56fe\u7ed3\u6784\u5bf9\u4e8e\u56fe\u6570\u636e\u6765\u8bf4\u66f4\u4e3a\u91cd\u8981\u3002\u53e6\u5916\uff0cGraphLime\u662f\u4e3a\u4e86\u89e3\u91ca\u8282\u70b9\u5206\u7c7b\u9884\u6d4b\u800c\u63d0\u51fa\u7684\uff0c\u4f46\u4e0d\u80fd\u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u5206\u7c7b\u6a21\u578b\u3002<\/span><\/section>\n

            2\uff09RelEx<\/span><\/h4>\n
            RelEx[57]\u7ed3\u5408\u4ee3\u7406\u65b9\u6cd5\u548c\u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\u7684\u601d\u60f3\uff0c\u7814\u7a76\u8282\u70b9\u5206\u7c7b\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u3002\u7ed9\u5b9a\u4e00\u4e2a\u76ee\u6807\u8282\u70b9\u53ca\u5176\u8ba1\u7b97\u56fe(N-hop\u90bb\u5c45)\uff0c\u5b83\u9996\u5148\u4ece\u8ba1\u7b97\u56fe\u4e2d\u968f\u673a\u91c7\u6837\u8fde\u63a5\u7684\u5b50\u56fe<\/strong>\uff0c\u83b7\u5f97\u4e00\u4e2a\u5c40\u90e8\u6570\u636e\u96c6\uff0c\u5e76\u5c06\u8fd9\u4e9b\u5b50\u56fe\u5582\u5165\u8bad\u7ec3\u597d\u7684GNN\uff0c\u4ee5\u83b7\u5f97\u5176\u9884\u6d4b\u7ed3\u679c\u3002\u4ece\u76ee\u6807\u8282\u70b9\u5f00\u59cb\uff0c\u5b83\u4ee5BFS\u7684\u65b9\u5f0f\u968f\u673a\u9009\u62e9\u76f8\u90bb\u8282\u70b9\u3002\u91c7\u7528GCN\u6a21\u578b\u4f5c\u4e3a\u4ee3\u7406\u6a21\u578b\u6765\u62df\u5408\u5c40\u90e8\u6570\u636e\u96c6\u3002\u4e0eGraphLime\u4e0d\u540c\uff0cRelEx\u4e2d\u7684\u4ee3\u7406\u6a21\u578b\u662f\u4e0d\u53ef\u89e3\u91ca\u7684\u3002\u8bad\u7ec3\u540e\uff0c\u5b83\u8fdb\u4e00\u6b65\u5e94\u7528\u524d\u8ff0\u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\uff0c\u5982\u751f\u6210\u8f6f\u63a9\u7801\u6216Gumbel-Softmax\u63a9\u7801\u6765\u89e3\u91ca\u9884\u6d4b\u7ed3\u679c\u3002\u8be5\u8fc7\u7a0b\u5305\u542b\u4e86\u591a\u4e2a\u6b65\u9aa4\u7684\u8fd1\u4f3c\uff0c\u6bd4\u5982\u4f7f\u7528\u4ee3\u7406\u6a21\u578b\u6765\u8fd1\u4f3c\u5c40\u90e8\u5173\u7cfb\uff0c\u4f7f\u7528\u63a9\u7801\u6765\u8fd1\u4f3c\u8fb9\u7684\u91cd\u8981\u6027\uff0c\u4ece\u800c\u4f7f\u5f97\u89e3\u91ca\u7684\u8bf4\u670d\u529b\u548c\u53ef\u4fe1\u5ea6\u964d\u4f4e\u3002\u7531\u4e8e\u53ef\u4ee5\u76f4\u63a5\u91c7\u7528\u57fa\u4e8e\u6270\u52a8\u7684\u65b9\u6cd5\u6765\u89e3\u91ca\u539f\u6709\u7684\u6df1\u5ea6\u56fe\u6a21\u578b\uff0c\u56e0\u6b64\u6ca1\u6709\u5fc5\u8981\u518d\u5efa\u7acb\u4e00\u4e2a\u4e0d\u53ef\u89e3\u91ca\u7684\u6df1\u5ea6\u6a21\u578b\u4f5c\u4e3a\u4ee3\u7406\u6a21\u578b\u6765\u89e3\u91ca\u3002<\/span><\/section>\n

            3\uff09PGM-Explainer<\/span><\/h4>\n
            PGM-Explainer[58]\u5efa\u7acb\u4e86\u4e00\u4e2a\u6982\u7387\u56fe\u5f62\u6a21\u578b<\/strong>\uff0c\u4e3aGNN\u63d0\u4f9b\u5b9e\u4f8b\u7ea7\u89e3\u91ca\u3002\u5c40\u90e8\u6570\u636e\u96c6\u662f\u901a\u8fc7\u968f\u673a\u8282\u70b9\u7279\u5f81\u6270\u52a8\u83b7\u5f97\u7684\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u56fe\uff0c\u6bcf\u6b21PGM-Explainer\u90fd\u4f1a\u968f\u673a\u6270\u52a8\u8ba1\u7b97\u56fe\u4e2d\u51e0\u4e2a\u968f\u673a\u8282\u70b9\u7684\u8282\u70b9\u7279\u5f81\u3002\u7136\u540e\u5bf9\u4e8e\u8ba1\u7b97\u56fe\u4e2d\u7684\u4efb\u4f55\u4e00\u4e2a\u8282\u70b9\uff0cPGM-Explainer\u90fd\u4f1a\u8bb0\u5f55\u4e00\u4e2a\u968f\u673a\u53d8\u91cf\uff0c\u8868\u793a\u5176\u7279\u5f81\u662f\u5426\u53d7\u5230\u6270\u52a8\uff0c\u4ee5\u53ca\u5176\u5bf9GNN\u9884\u6d4b\u7684\u5f71\u54cd\u3002\u901a\u8fc7\u591a\u6b21\u91cd\u590d\u8fd9\u6837\u7684\u8fc7\u7a0b\uff0c\u5c31\u53ef\u4ee5\u5f97\u5230\u4e00\u4e2a\u5c40\u90e8\u6570\u636e\u96c6\u3002\u901a\u8fc7Grow-Shrink (GS) \u7b97\u6cd5[66]\u9009\u62e9\u4f9d\u8d56\u6027\u6700\u5f3a\u7684\u53d8\u91cf\u6765\u51cf\u5c0f\u5c40\u90e8\u6570\u636e\u96c6\u7684\u5927\u5c0f\u3002\u6700\u540e\u91c7\u7528\u53ef\u89e3\u91ca\u7684\u8d1d\u53f6\u65af\u7f51\u7edc\u6765\u62df\u5408\u5c40\u90e8\u6570\u636e\u96c6<\/strong>\uff0c\u5e76\u89e3\u91ca\u539f\u59cbGNN\u6a21\u578b\u7684\u9884\u6d4b\u3002PGM-Explainer\u53ef\u4ee5\u63d0\u4f9b\u6709\u5173\u56fe\u8282\u70b9\u7684\u89e3\u91ca\uff0c\u4f46\u5ffd\u7565\u4e86\u5305\u542b\u91cd\u8981\u56fe\u62d3\u6251\u4fe1\u606f\u7684\u56fe\u8fb9\u3002\u6b64\u5916\uff0c\u4e0eGraphLime\u548cRelEx\u4e0d\u540c\u7684\u662f\uff0cPGM-Explainer\u53ef\u4ee5\u540c\u65f6\u7528\u4e8e\u89e3\u91ca\u8282\u70b9\u5206\u7c7b\u548c\u56fe\u5f62\u5206\u7c7b\u4efb\u52a1\u3002<\/span><\/section>\n

            3.4 \u5206\u89e3\u65b9\u6cd5\uff08Decomposition Methods\uff09<\/span><\/h3>\n
            \u5206\u89e3\u65b9\u6cd5\u662f\u53e6\u4e00\u79cd\u6bd4\u8f83\u6d41\u884c\u7684\u89e3\u91ca\u6df1\u5ea6\u56fe\u50cf\u5206\u7c7b\u5668\u7684\u65b9\u6cd5\uff0c\u5b83\u901a\u8fc7\u5c06\u539f\u59cb\u6a21\u578b\u9884\u6d4b\u5206\u89e3\u4e3a\u82e5\u5e72\u9879\u6765\u8861\u91cf\u8f93\u5165\u7279\u5f81\u7684\u91cd\u8981\u6027<\/strong>\u3002\u7136\u540e\u5c06\u8fd9\u4e9b\u9879\u89c6\u4e3a\u76f8\u5e94\u8f93\u5165\u7279\u5f81\u7684\u91cd\u8981\u6027\u5206\u6570\u3002\u8fd9\u4e9b\u65b9\u6cd5\u76f4\u63a5\u7814\u7a76\u6a21\u578b\u53c2\u6570\u6765\u63ed\u793a\u8f93\u5165\u7a7a\u95f4\u4e2d\u7684\u7279\u5f81\u4e0e\u8f93\u51fa\u9884\u6d4b\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u8981\u6c42\u5206\u89e3\u9879\u4e4b\u548c\u7b49\u4e8e\u539f\u59cb\u9884\u6d4b\u5f97\u5206\u3002\u7531\u4e8e\u56fe\u5305\u542b\u8282\u70b9\u3001\u8fb9\u548c\u8282\u70b9\u7279\u5f81\uff0c\u56e0\u6b64\u5c06\u8fd9\u7c7b\u65b9\u6cd5\u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u57df\u662f\u5177\u6709\u6311\u6218\u6027\u7684\u3002\u5f88\u96be\u5c06\u5206\u6570\u5206\u914d\u7ed9\u4e0d\u540c\u7684\u8fb9\uff0c\u56fe\u6570\u636e\u8fb9\u5305\u542b\u7740\u91cd\u8981\u7684\u7ed3\u6784\u4fe1\u606f\uff0c\u4e0d\u5bb9\u5ffd\u89c6\u3002<\/span><\/section>\n
            <\/div>\n
            \u56fe5 \u5206\u89e3\u65b9\u6cd5\u7684\u4e00\u822c\u6d41\u7a0b<\/span><\/section>\n
            \u672c\u6587\u5c06\u4ecb\u7ecd\u6700\u8fd1\u63d0\u51fa\u7684\u51e0\u79cd\u7528\u4e8e\u89e3\u91ca\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u6cdb\u8bfb\u5206\u89e3\u65b9\u6cd5\uff0c\u5305\u62ec\uff1aLayerwise Relevance Propagation\uff08LRP\uff09[49]\u3001[54]\u3001Excitation BP[50]\u548cGNN-LRP[55]\u3002\u8fd9\u4e9b\u7b97\u6cd5\u7684\u4e3b\u8981\u601d\u60f3\u662f\u5efa\u7acb\u5206\u6570\u5206\u89e3\u89c4\u5219\uff0c\u5c06\u9884\u6d4b\u5206\u6570\u5206\u914d\u5230\u8f93\u5165\u7a7a\u95f4\u3002<\/strong>\u8fd9\u4e9b\u65b9\u6cd5\u7684\u4e00\u822c\u6d41\u7a0b\u5982\u56fe4\u6240\u793a\u3002\u4ee5\u53cd\u5411\u4f20\u64ad\u7684\u65b9\u5f0f\u9010\u5c42\u5206\u53d1\u9884\u6d4b\u5f97\u5206\uff0c\u76f4\u5230\u8f93\u5165\u5c42\u3002<\/strong>\u4ece\u8f93\u51fa\u5c42\u5f00\u59cb\uff0c\u5c06\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u4f5c\u4e3a\u521d\u59cb\u76ee\u6807\u5206\u6570\u3002\u7136\u540e\u5c06\u5206\u6570\u8fdb\u884c\u5206\u89e3\uff0c\u5e76\u6309\u7167\u5206\u89e3\u89c4\u5219\u5206\u914d\u7ed9\u4e0a\u4e00\u5c42\u7684\u795e\u7ecf\u5143\u3002\u901a\u8fc7\u91cd\u590d\u8fd9\u6837\u7684\u8fc7\u7a0b\uff0c\u76f4\u5230\u8f93\u5165\u7a7a\u95f4\uff0c\u5b83\u4eec\u53ef\u4ee5\u5f97\u5230\u8282\u70b9\u7279\u5f81\u7684\u91cd\u8981\u6027\u5206\u6570<\/strong>\uff0c\u8fd9\u4e9b\u5206\u6570\u53ef\u4ee5\u7ec4\u5408\u8d77\u6765\u8868\u793a\u8fb9\u91cd\u8981\u6027\u3001\u8282\u70b9\u91cd\u8981\u6027\u548c\u6e38\u8d70\u91cd\u8981\u6027\u3002\u4f46\u662f\u8fd9\u4e9b\u7b97\u6cd5\u90fd\u5ffd\u7565\u4e86\u6df1\u5ea6\u56fe\u6a21\u578b\u4e2d\u7684\u6fc0\u6d3b\u51fd\u6570\u3002\u4e0d\u540c\u5206\u89e3\u65b9\u6cd5\u7684\u4e3b\u8981\u533a\u522b\u5728\u4e8e\u5206\u6570\u5206\u89e3\u89c4\u5219\u548c\u89e3\u91ca\u7684\u76ee\u6807\u3002<\/strong><\/span><\/section>\n

            1\uff09LRP<\/strong><\/span><\/h4>\n
            LRP[49]\uff0c[54]\u5c06\u539f\u6765\u7684LRP\u7b97\u6cd5[67]\u6269\u5c55\u5230\u6df1\u5ea6\u56fe\u6a21\u578b\u3002\u5b83\u5c06\u8f93\u51fa\u7684\u9884\u6d4b\u5206\u6570\u5206\u89e3\u4e3a\u4e0d\u540c\u7684\u8282\u70b9\u91cd\u8981\u6027\u5206\u6570<\/strong>\u3002\u5206\u6570\u5206\u89e3\u89c4\u5219\u662f\u57fa\u4e8e\u9690\u85cf\u7279\u5f81\u548c\u6743\u91cd\u5236\u5b9a\u7684\u3002<\/strong>\u5bf9\u4e8e\u4e00\u4e2a\u76ee\u6807\u795e\u7ecf\u5143\uff0c\u5176\u5f97\u5206\u8868\u793a\u4e3a\u4e0a\u4e00\u5c42\u795e\u7ecf\u5143\u5f97\u5206\u7684\u7ebf\u6027\u8fd1\u4f3c\u3002\u76ee\u6807\u795e\u7ecf\u5143\u6fc0\u6d3b\u8d21\u732e\u5ea6\u8f83\u9ad8\u7684\u795e\u7ecf\u5143\u83b7\u5f97\u7684\u76ee\u6807\u795e\u7ecf\u5143\u5f97\u5206\u6bd4\u4f8b\u8f83\u5927\u3002\u4e3a\u4e86\u6ee1\u8db3\u4fdd\u5b88\u5c5e\u6027\uff0c\u5728\u4e8b\u540e\u89e3\u91ca\u9636\u6bb5\u5c06\u90bb\u63a5\u77e9\u9635\u4f5c\u4e3aGNN\u6a21\u578b\u7684\u4e00\u90e8\u5206\uff0c\u8fd9\u6837\u5728\u5206\u6570\u5206\u914d\u65f6\u5c31\u53ef\u4ee5\u5ffd\u7565\u5b83\uff0c\u5426\u5219\uff0c\u90bb\u63a5\u77e9\u9635\u4e5f\u4f1a\u6536\u5230\u5206\u89e3\u540e\u7684\u5206\u6570\uff0c\u4ece\u800c\u4f7f\u4fdd\u5b88\u5c5e\u6027\u5931\u6548\u3002\u7531\u4e8eLRP\u662f\u76f4\u63a5\u6839\u636e\u6a21\u578b\u53c2\u6570\u5f00\u53d1\u7684\uff0c\u6240\u4ee5\u5176\u89e3\u91ca\u7ed3\u679c\u66f4\u5177\u6709\u53ef\u4fe1\u5ea6\u3002\u4f46\u5b83\u53ea\u80fd\u7814\u7a76\u4e0d\u540c\u8282\u70b9\u7684\u91cd\u8981\u6027\uff0c\u4e0d\u80fd\u5e94\u7528\u4e8e\u56fe\u7ed3\u6784\uff0c\u5982\u5b50\u56fe\u548c\u6e38\u8d70\u3002\u8be5\u7b97\u6cd5\u9700\u8981\u5bf9\u6a21\u578b\u7ed3\u6784\u6709\u5168\u9762\u7684\u4e86\u89e3\uff0c\u8fd9\u5c31\u9650\u5236\u4e86\u5b83\u5bf9\u975e\u4e13\u4e1a\u7528\u6237\u7684\u5e94\u7528\uff0c\u5982\u8de8\u5b66\u79d1\u7814\u7a76\u4eba\u5458\u3002<\/span><\/section>\n

            2\uff09Excitation BP<\/span><\/h4>\n
            Excitation BP[50]\u4e0eLRP\u7b97\u6cd5\u6709\u7740\u76f8\u4f3c\u7684\u601d\u60f3\uff0c\u4f46\u5b83\u662f\u57fa\u4e8e\u5168\u6982\u7387\u6cd5\u5219\u5f00\u53d1\u7684\u3002\u5b83\u5b9a\u4e49\u4e86\u5f53\u524d\u5c42\u4e2d\u4e00\u4e2a\u795e\u7ecf\u5143\u7684\u6982\u7387\u7b49\u4e8e\u5b83\u8f93\u51fa\u7ed9\u4e0b\u4e00\u5c42\u6240\u6709\u8fde\u63a5\u795e\u7ecf\u5143\u7684\u603b\u6982\u7387\u3002\u90a3\u4e48\u5206\u6570\u5206\u89e3\u89c4\u5219\u53ef\u4ee5\u770b\u4f5c\u662f\u5c06\u76ee\u6807\u6982\u7387\u5206\u89e3\u4e3a\u51e0\u4e2a\u6761\u4ef6\u6982\u7387\u9879\u3002Excitation BP\u7684\u8ba1\u7b97\u4e0eLRP\u4e2d\u7684z+\u89c4\u5219\u9ad8\u5ea6\u76f8\u4f3c\u3002\u56e0\u6b64\u5b83\u4e0eLRP\u7b97\u6cd5\u6709\u7740\u76f8\u540c\u7684\u4f18\u70b9\u548c\u5c40\u9650\u6027\u3002<\/span><\/section>\n

            3\uff09GNN-LRP<\/span><\/h4>\n
            GNN-LRP[55]\u7814\u7a76\u4e86\u4e0d\u540c\u56fe\u6e38\u8d70<\/strong>\u7684\u91cd\u8981\u6027\u3002\u7531\u4e8e\u5728\u8fdb\u884c\u90bb\u57df\u4fe1\u606f\u805a\u5408\u65f6\uff0c\u56fe\u6e38\u8d70\u5bf9\u5e94\u4e8e\u6d88\u606f\u6d41<\/strong>\uff0c\u56e0\u6b64\u5b83\u5bf9\u6df1\u5c42\u56fe\u795e\u7ecf\u7f51\u7edc\u66f4\u5177\u6709\u4e00\u81f4\u6027\u3002\u5f97\u5206\u5206\u89e3\u89c4\u5219\u662f\u6a21\u578b\u9884\u6d4b\u7684\u9ad8\u9636\u6cf0\u52d2\u5206\u89e3\u3002\u7814\u7a76\u8868\u660e\uff0c\u6cf0\u52d2\u5206\u89e3<\/span>\uff08\u5728\u6839\u96f6\u5904\uff09<\/span>\u53ea\u5305\u542bT\u9636\u9879\uff0c\u5176\u4e2dT\u662f\u8bad\u7ec3\u7684GNN\u7684\u5c42\u6570\u3002\u90a3\u4e48\u6bcf\u4e2a\u9879\u5bf9\u5e94\u4e00\u4e2aT\u9636\u56fe\u6e38\u8d70\uff0c\u53ef\u4ee5\u89c6\u4e3a\u5176\u91cd\u8981\u6027\u5f97\u5206\u3002\u7531\u4e8e\u65e0\u6cd5\u76f4\u63a5\u8ba1\u7b97\u6cf0\u52d2\u5c55\u5f00\u7ed9\u51fa\u7684\u9ad8\u9636\u5bfc\u6570\uff0cGNN-LRP\u8fd8\u9075\u5faa\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u6765\u903c\u8fd1T\u9636\u9879\u3002GNN-LRP\u4e2d\u7684\u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\u4e0eLRP\u7b97\u6cd5\u7c7b\u4f3c\u3002\u7136\u800c\uff0cGNN-LRP\u4e0d\u662f\u5c06\u5206\u6570\u5206\u914d\u7ed9\u8282\u70b9\u6216\u8fb9\uff0c\u800c\u662f\u5c06\u5206\u6570\u5206\u914d\u7ed9\u4e0d\u540c\u7684\u56fe\u6e38\u8d70\u3002\u5b83\u8bb0\u5f55\u4e86\u5c42\u4e0e\u5c42\u4e4b\u95f4\u7684\u6d88\u606f\u5206\u53d1\u8fc7\u7a0b\u7684\u8def\u5f84<\/strong>\u3002\u8fd9\u4e9b\u8def\u5f84\u88ab\u8ba4\u4e3a\u662f\u4e0d\u540c\u7684\u6e38\u8d70\uff0c\u5e76\u4ece\u5b83\u4eec\u5bf9\u5e94\u7684\u8282\u70b9\u4e0a\u83b7\u5f97\u5206\u6570\u3002\u867d\u7136GNN-LRP\u5177\u6709\u575a\u5b9e\u7684\u7406\u8bba\u80cc\u666f\uff0c\u4f46\u5176\u8ba1\u7b97\u4e2d\u7684\u8fd1\u4f3c\u503c\u53ef\u80fd\u5e76\u4e0d\u51c6\u786e\u3002\u7531\u4e8e\u6bcf\u4e2a\u6e38\u8d70\u90fd\u8981\u5355\u72ec\u8003\u8651\uff0c\u8ba1\u7b97\u590d\u6742\u5ea6\u5f88\u9ad8\u3002\u6b64\u5916\uff0c\u5bf9\u4e8e\u975e\u4e13\u4e1a\u4eba\u5458\u6765\u8bf4\uff0c\u5b83\u7684\u4f7f\u7528\u4e5f\u5177\u6709\u6311\u6218\u6027\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u8de8\u5b66\u79d1\u9886\u57df\u3002<\/span><\/section>\n
            <\/p>\n
            \n

             <\/span><\/p>\n<\/section>\n

            <\/span><\/section>\n

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

            <\/section>\n
            \n
            <\/svg><\/section>\n<\/section>\n<\/section>\n
            \n

            4. \u6a21\u578b\u7ea7\u65b9\u6cd5<\/strong><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

            <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
            \u4e0e\u5b9e\u4f8b\u7ea7\u65b9\u6cd5\u4e0d\u540c\uff0c\u6a21\u578b\u7ea7\u65b9\u6cd5\u65e8\u5728\u63d0\u4f9b\u4e00\u822c\u6027\u7684\u89c1\u89e3\u548c\u9ad8\u5c42\u6b21\u7684\u7406\u89e3\u6765\u89e3\u91ca\u6df1\u5c42\u56fe\u6a21\u578b\u3002\u5b83\u4eec\u7814\u7a76\u4ec0\u4e48\u6837\u7684\u8f93\u5165\u56fe\u6a21\u5f0f\u53ef\u4ee5\u5bfc\u81f4GNN\u7684\u67d0\u79cd\u884c\u4e3a<\/strong>\uff0c\u4f8b\u5982\u6700\u5927\u5316\u76ee\u6807\u9884\u6d4b\u3002\u8f93\u5165\u4f18\u5316[16]\u662f\u83b7\u5f97\u56fe\u50cf\u5206\u7c7b\u5668\u6a21\u578b\u7ea7\u89e3\u91ca\u7684\u4e00\u4e2a\u70ed\u95e8\u65b9\u5411\u3002\u4f46\u662f\uff0c\u7531\u4e8e\u56fe\u62d3\u6251\u4fe1\u606f\u7684\u79bb\u6563\u6027\uff0c\u5b83\u4e0d\u80fd\u76f4\u63a5\u5e94\u7528\u4e8e\u56fe\u6a21\u578b\uff0c\u4ece\u800c\u4f7fGNN\u5728\u6a21\u578b\u5c42\u9762\u7684\u89e3\u91ca\u66f4\u5177\u6311\u6218\u6027\u3002\u5b83\u4ecd\u7136\u662f\u4e00\u4e2a\u91cd\u8981\u4f46\u7814\u7a76\u8f83\u5c11\u7684\u8bfe\u9898\u3002\u636e\u6211\u4eec\u6240\u77e5\uff0c\u73b0\u6709\u7684\u89e3\u91ca\u56fe\u795e\u7ecf\u7f51\u7edc\u7684\u6a21\u578b\u7ea7\u65b9\u6cd5\u53ea\u6709XGNN[41]\u3002<\/span><\/section>\n

            1\uff09XGNN<\/span><\/h4>\n
            XGNN[41]\u63d0\u51fa\u901a\u8fc7\u56fe\u751f\u6210\u6765\u89e3\u91caGNN\u3002\u5b83\u4e0d\u662f\u76f4\u63a5\u4f18\u5316\u8f93\u5165\u56fe\uff0c\u800c\u662f\u8bad\u7ec3\u4e00\u4e2a\u56fe\u751f\u6210\u5668\uff0c\u4f7f\u751f\u6210\u7684\u56fe\u80fd\u591f\u6700\u5927\u5316\u76ee\u6807\u56fe\u9884\u6d4b\u3002\u7136\u540e\uff0c\u751f\u6210\u7684\u56fe\u88ab\u89c6\u4e3a\u76ee\u6807\u9884\u6d4b\u7684\u89e3\u91ca\uff0c\u5e76\u88ab\u671f\u671b\u5305\u542b\u5224\u522b\u6027\u7684\u56fe\u6a21\u5f0f\u3002\u5728XGNN\u4e2d\uff0c\u56fe\u5f62\u751f\u6210\u88ab\u8868\u8ff0\u4e3a\u4e00\u4e2a\u5f3a\u5316\u5b66\u4e60\u95ee\u9898\u3002\u5bf9\u4e8e\u6bcf\u4e00\u6b65\uff0c\u751f\u6210\u5668\u90fd\u4f1a\u9884\u6d4b\u5982\u4f55\u5728\u5f53\u524d\u56fe\u4e2d\u589e\u52a0\u4e00\u6761\u8fb9\u3002\u7136\u540e\u5c06\u751f\u6210\u7684\u56fe\u8f93\u5165\u5230\u8bad\u7ec3\u597d\u7684GNN\u4e2d\uff0c\u901a\u8fc7\u7b56\u7565\u68af\u5ea6\u83b7\u5f97\u53cd\u9988\u6765\u8bad\u7ec3\u751f\u6210\u5668\u3002\u6b64\u5916\uff0c\u8fd8\u52a0\u5165\u4e86\u4e00\u4e9b\u56fe\u89c4\u5219\uff0c\u4ee5\u9f13\u52b1\u89e3\u91ca\u65e2\u6709\u6548\u53c8\u80fd\u88ab\u4eba\u7c7b\u7406\u89e3\u3002XGNN\u662f\u4e00\u4e2a\u751f\u6210\u6a21\u578b\u7ea7\u89e3\u91ca\u7684\u901a\u7528\u6846\u67b6\uff0c\u56e0\u6b64\u53ef\u4ee5\u5e94\u7528\u4efb\u4f55\u5408\u9002\u7684\u56fe\u751f\u6210\u7b97\u6cd5\u3002\u8be5\u89e3\u91ca\u662f\u901a\u7528\u7684\uff0c\u5e76\u4e14\u63d0\u4f9b\u4e86\u5bf9\u8bad\u7ec3\u7684GNNs\u7684\u5168\u5c40\u7406\u89e3\u3002\u7136\u800cXGNN\u53ea\u8bc1\u660e\u4e86\u5176\u5728\u89e3\u91ca\u56fe\u5206\u7c7b\u6a21\u578b\u65b9\u9762\u7684\u6709\u6548\u6027\uff0cXGNN\u662f\u5426\u53ef\u4ee5\u5e94\u7528\u4e8e\u8282\u70b9\u5206\u7c7b\u4efb\u52a1\u8fd8\u4e0d\u5f97\u800c\u77e5\uff0c\u8fd9\u662f\u672a\u6765\u7814\u7a76\u4e2d\u9700\u8981\u63a2\u7d22\u7684\u91cd\u8981\u65b9\u5411\u3002<\/span><\/section>\n
            <\/p>\n
            \n

             <\/span><\/p>\n<\/section>\n

            <\/span><\/section>\n

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

            <\/section>\n
            \n
            <\/svg><\/section>\n<\/section>\n<\/section>\n
            \n

            5. \u8bc4\u4f30\u6a21\u578b<\/strong><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

            <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
            \u7531\u4e8e\u7f3a\u4e4f ground truths\uff0c\u56e0\u6b64\u4e0d\u5bb9\u6613\u5bf9\u89e3\u91ca\u65b9\u6cd5\u7684\u7ed3\u679c\u8fdb\u884c\u8bc4\u4f30\uff0c\u4f5c\u8005\u8ba8\u8bba\u5e76\u5206\u6790\u4e86\u51e0\u79cd\u5e38\u7528\u7684\u6570\u636e\u96c6\u548c\u5ea6\u91cf\u6807\u51c6\u3002<\/span><\/section>\n

            5.1. Datasets<\/span><\/h3>\n
            \u9700\u8981\u9009\u62e9\u5408\u9002\u7684\u6570\u636e\u96c6\u6765\u8bc4\u4f30\u4e0d\u540c\u7684\u89e3\u91ca\u6280\u672f\uff0c\u5e76\u4e14\u5e0c\u671b\u6570\u636e\u662f\u76f4\u89c2\u7684\uff0c\u6613\u4e8e\u53ef\u89c6\u5316\u7684\u3002\u5e94\u8be5\u5728\u6570\u636e\u5b9e\u4f8b\u548c\u6807\u7b7e\u4e4b\u95f4\u8574\u542b\u4eba\u7c7b\u53ef\u4ee5\u7406\u89e3\u7684\u7406\u7531\uff0c\u8fd9\u6837\u4e13\u5bb6\u5c31\u53ef\u4ee5\u9a8c\u8bc1\u8fd9\u4e9b\u7406\u7531\u662f\u5426\u88ab\u89e3\u91ca\u7b97\u6cd5\u8bc6\u522b<\/strong>\u3002\u4e3a\u4e86\u8bc4\u4f30\u4e0d\u540c\u7684\u89e3\u91ca\u6280\u672f\uff0c\u901a\u5e38\u91c7\u7528\u51e0\u79cd\u7c7b\u578b\u7684\u6570\u636e\u96c6\uff0c\u5305\u62ec\u5408\u6210\u6570\u636e\u3001\u60c5\u611f\u56fe\u6570\u636e\u548c\u5206\u5b50\u6570\u636e\u3002<\/span><\/section>\n

            5.1.1. Synthetic data<\/span><\/h3>\n
            \u5229\u7528\u73b0\u6709\u7684\u5408\u6210\u6570\u636e\u96c6\u6765\u8bc4\u4f30\u89e3\u91ca\u6280\u672f[42]\uff0c[43]\u3002\u5728\u8fd9\u6837\u7684\u6570\u636e\u96c6\u4e2d\uff0c\u5305\u542b\u4e86\u4e0d\u540c\u7684 graph motifs\uff0c\u53ef\u4ee5\u901a\u8fc7\u5b83\u4eec\u786e\u5b9a\u8282\u70b9\u6216\u56fe\u7684\u6807\u7b7e\u3002\u6570\u636e\u5b9e\u4f8b\u548c\u6570\u636e\u6807\u7b7e\u4e4b\u95f4\u7684\u5173\u7cfb\u7531\u4eba\u7c7b\u5b9a\u4e49\u3002\u5373\u4f7f\u7ecf\u8fc7\u8bad\u7ec3\u7684GNNs\u53ef\u80fd\u65e0\u6cd5\u5b8c\u7f8e\u5730\u6355\u6349\u5230\u8fd9\u6837\u7684\u5173\u7cfb\uff0c\u4f46graph motifs \u53ef\u4ee5\u4f5c\u4e3a\u89e3\u91ca\u7ed3\u679c\u7684ground truths \u7684\u5408\u7406\u8fd1\u4f3c\u503c\u3002\u8fd9\u91cc\u6211\u4eec\u4ecb\u7ecd\u51e0\u79cd\u5e38\u89c1\u7684\u5408\u6210\u6570\u636e\u96c6\u3002<\/span><\/section>\n
            BA-shapes\uff1a<\/strong>\u5b83\u662f\u4e00\u4e2a\u8282\u70b9\u5206\u7c7b\u6570\u636e\u96c6\uff0c\u67094\u4e2a\u4e0d\u540c\u7684\u8282\u70b9\u6807\u7b7e\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u56fe\u5f62\uff0c\u5b83\u5305\u542b\u4e00\u4e2a\u57fa\u7840\u56fe<\/span>\uff08300\u4e2a\u8282\u70b9\uff09<\/span>\u548c\u4e00\u4e2a\u7c7b\u4f3c\u623f\u5b50\u76845\u8282\u70b9 motif\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u57fa\u7840\u56fe\u662f\u7531Barab\u00b4asi-Albert\uff08BA\uff09\u6a21\u578b\u83b7\u5f97\u7684\uff0c\u5b83\u53ef\u4ee5\u751f\u6210\u5177\u6709\u4f18\u5148\u9644\u52a0\u673a\u5236\u7684\u968f\u673a\u65e0\u6807\u5ea6\u7f51\u7edc[68]\u3002motif \u88ab\u9644\u52a0\u5230\u57fa\u56fe\u4e0a\uff0c\u540c\u65f6\u6dfb\u52a0\u968f\u673a\u8fb9\u3002\u6bcf\u4e2a\u8282\u70b9\u6839\u636e\u5176\u662f\u5426\u5c5e\u4e8e\u57fa\u7840\u56fe\u6216motif \u7684\u4e0d\u540c\u7a7a\u95f4\u4f4d\u7f6e\u8fdb\u884c\u6807\u6ce8\u3002<\/span><\/section>\n
            BA-Community\uff1a<\/strong>\u8fd9\u662f\u4e00\u4e2a\u67098\u4e2a\u4e0d\u540c\u6807\u7b7e\u7684\u8282\u70b9\u5206\u7c7b\u6570\u636e\u96c6\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u56fe\uff0c\u5b83\u662f\u901a\u8fc7\u7ec4\u5408\u4e24\u4e2a\u968f\u673a\u6dfb\u52a0\u8fb9\u7684BA-shapes\u56fe\u83b7\u5f97\u7684\u3002\u8282\u70b9\u6807\u7b7e\u7531BA-shapes\u56fe\u7684\u6210\u5458\u8d44\u683c\u53ca\u5176\u7ed3\u6784\u4f4d\u7f6e\u51b3\u5b9a\u3002<\/span><\/section>\n
            Tree-Cycle\uff1a<\/strong>\u5b83\u662f\u4e00\u4e2a\u6709\u4e24\u4e2a\u4e0d\u540c\u6807\u7b7e\u7684\u8282\u70b9\u5206\u7c7b\u6570\u636e\u96c6\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u56fe\uff0c\u5b83\u7531\u6df1\u5ea6\u7b49\u4e8e8\u7684\u57fa\u5e73\u8861\u6811\u56fe\u548c6\u8282\u70b9\u5468\u671f motif \u7ec4\u6210\u3002\u8fd9\u4e24\u90e8\u5206\u662f\u968f\u673a\u8fde\u63a5\u7684\u3002\u57fa\u56fe\u4e2d\u8282\u70b9\u7684\u6807\u7b7e\u4e3a0\uff0c\u5426\u5219\u4e3a1\u3002<\/span><\/section>\n
            Tree-Grids\uff1a<\/strong>\u5b83\u662f\u4e00\u4e2a\u6709\u4e24\u4e2a\u4e0d\u540c\u6807\u7b7e\u7684\u8282\u70b9\u5206\u7c7b\u6570\u636e\u96c6\u3002\u5b83\u4e0e Tree-Cycle \u6570\u636e\u96c6\u76f8\u540c\uff0c\u53ea\u662fTree-Grids\u6570\u636e\u96c6\u91c7\u7528\u4e869\u8282\u70b9\u7f51\u683c motifs \u800c\u4e0d\u662f\u5468\u671f motifs \u3002<\/span><\/section>\n
            BA-2Motifs\uff1a<\/strong>\u5b83\u662f\u4e00\u4e2a\u5177\u67092\u79cd\u4e0d\u540c\u56fe\u6807\u7b7e\u7684\u56fe\u5f62\u5206\u7c7b\u6570\u636e\u96c6\u3002\u6709800\u4e2a\u56fe\uff0c\u6bcf\u4e2a\u56fe\u90fd\u662f\u901a\u8fc7\u5728\u57fa\u7840BA\u56fe\u4e0a\u9644\u52a0\u4e0d\u540c\u7684motif\u6765\u83b7\u5f97\u7684\uff0c\u5982house-like motif \u548c five-node cycle motif\u3002\u4e0d\u540c\u7684\u56fe\u662f\u6839\u636emotif \u7684\u7c7b\u578b\u6765\u6807\u6ce8\u7684\u3002<\/span><\/section>\n
            \u5728\u8fd9\u4e9b\u6570\u636e\u96c6\u4e2d\uff0c\u6240\u6709\u8282\u70b9\u7279\u5f81\u90fd\u88ab\u521d\u59cb\u5316\u4e3a\u51681\u5411\u91cf\u3002\u8bad\u7ec3\u597d\u7684GNNs\u6a21\u578b\u8981\u6355\u6349\u56fe\u7ed3\u6784\u6765\u8fdb\u884c\u9884\u6d4b\u3002\u7136\u540e\u6839\u636e\u6bcf\u4e2a\u6570\u636e\u96c6\u7684\u6784\u5efa\u89c4\u5219\uff0c\u6211\u4eec\u53ef\u4ee5\u5206\u6790\u89e3\u91ca\u7ed3\u679c\u3002\u4f8b\u5982\uff0c\u5728BA-2Motifs\u6570\u636e\u96c6\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u7814\u7a76\u89e3\u91ca\u662f\u5426\u80fd\u591f\u6355\u83b7motif\u7ed3\u6784\u3002\u7136\u800c\uff0c\u5408\u6210\u6570\u636e\u96c6\u53ea\u5305\u542b\u56fe\u548c\u6807\u7b7e\u4e4b\u95f4\u7684\u7b80\u5355\u5173\u7cfb\uff0c\u53ef\u80fd\u4e0d\u8db3\u4ee5\u8fdb\u884c\u7efc\u5408\u8bc4\u4f30\u3002<\/span><\/section>\n

            5.1.2 Sentiment graph data<\/span><\/h3>\n
            \u7531\u4e8e\u4eba\u7c7b\u53ea\u6709\u6709\u9650\u7684\u9886\u57df\u77e5\u8bc6\uff0c\u4f20\u7edf\u7684\u56fe\u6570\u636e\u96c6\u5728\u7406\u89e3\u4e0a\u5177\u6709\u6311\u6218\u6027\uff0c\u56e0\u6b64\u9700\u8981\u6784\u5efa\u4eba\u7c7b\u53ef\u7406\u89e3\u7684\u56fe\u6570\u636e\u96c6\u3002\u6587\u672c\u6570\u636e\u5177\u6709\u4eba\u7c7b\u53ef\u7406\u89e3\u7684\u8bed\u4e49\u7684\u5355\u8bcd\u548c\u77ed\u8bed\u7ec4\u6210\uff0c\u56e0\u6b64\u53ef\u4ee5\u6210\u4e3a\u56fe\u89e3\u91ca\u4efb\u52a1\u7684\u5408\u9002\u9009\u62e9\uff0c\u89e3\u91ca\u7ed3\u679c\u53ef\u4ee5\u5f88\u5bb9\u6613\u5730\u88ab\u4eba\u7c7b\u8bc4\u4f30\u3002\u56e0\u6b64\u6211\u4eec\u57fa\u4e8e\u6587\u672c\u60c5\u611f\u5206\u6790\u6570\u636e\u6784\u5efa\u4e86\u4e09\u4e2a\u60c5\u611f\u56fe\u6570\u636e\u96c6\uff0c\u5305\u62ecSST2[69]\u3001SST5[69]\u548cTwitter[70]\u6570\u636e\u96c6\u3002<\/span><\/section>\n
            \n
            fig7<\/span><\/figcaption><\/figure>\n<\/section>\n
            \t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\u56fe5 \u6587\u672c\u60c5\u611f\u56fe<\/span><\/section>\n
            \u5bf9\u4e8e\u6bcf\u4e2a\u6587\u672c\u5e8f\u5217\uff0c\u5c06\u5176\u8f6c\u6362\u4e3a\u4e00\u4e2a\u56fe\uff0c\u6bcf\u4e2a\u8282\u70b9\u4ee3\u8868\u4e00\u4e2a\u5355\u8bcd\uff0c\u800c\u8fb9\u5219\u53cd\u6620\u4e0d\u540c\u5355\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u4f5c\u8005\u91c7\u7528Biaffine\u89e3\u6790\u5668[71]\u6765\u63d0\u53d6\u8bcd\u7684\u4f9d\u8d56\u5173\u7cfb\u3002\u56fe5\u4e2d\u5c55\u793a\u4e86\u751f\u6210\u7684\u60c5\u611f\u56fe\u7684\u4e00\u4e2a\u4f8b\u5b50\u3002\u751f\u6210\u7684\u56fe\u662f\u6709\u5411\u7684\uff0c\u4f46\u8fb9\u6807\u7b7e\u88ab\u5ffd\u7565\u4e86\uff0c\u56e0\u4e3a\u5927\u591a\u6570GNNs\u4e0d\u80fd\u6355\u83b7\u8fb9\u6807\u7b7e\u4fe1\u606f\u3002\u7528BERT[72]\u6765\u5b66\u4e60\u5355\u8bcd\u5d4c\u5165\uff0c\u5e76\u5c06\u8fd9\u79cd\u5d4c\u5165\u4f5c\u4e3a\u56fe\u8282\u70b9\u7684\u521d\u59cb\u5d4c\u5165\u3002\u5efa\u7acb\u4e00\u4e2a\u6a21\u578b\uff0c\u91c7\u7528\u9884\u8bad\u7ec3\u597d\u7684BERT\u4f5c\u4e3a\u7279\u5f81\u63d0\u53d6\u5668\uff0c\u91c7\u7528\u4e00\u5c42\u5e73\u5747\u6c60\u5316\u7684GCN\u4f5c\u4e3a\u5206\u7c7b\u5668\u3002\u6700\u540e\u9884\u8bad\u7ec3\u7684BERT\u4e3a\u6bcf\u4e2a\u8bcd\u63d0\u53d6768\u7ef4\u7684\u7279\u5f81\u5411\u91cf\uff0c\u4f5c\u4e3a\u60c5\u611f\u56fe\u6570\u636e\u4e2d\u7684\u8282\u70b9\u7279\u5f81\u3002<\/span><\/section>\n
            <\/div>\n
            \u4f5c\u8005\u5efa\u7acb\u4e86\u4e09\u4e2a\u60c5\u611f\u56fe\u6570\u636e\u96c6\uff0c\u5206\u522b\u4e3aGraph-SST2\u3001Graph-SST5\u548cGraph-Twitter \uff0c\u5e76\u5373\u5c06\u516c\u5f00\uff0c\u53ef\u4ee5\u76f4\u63a5\u7528\u4e8e\u7814\u7a76\u4e0d\u540c\u7684\u53ef\u89e3\u91ca\u6280\u672f\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u7684\u7edf\u8ba1\u548c\u5c5e\u6027\u5982\u88682\u6240\u793a\u3002\u4e3a\u4e86\u9a8c\u8bc1\u672c\u6587\u751f\u6210\u7684\u60c5\u611f\u6570\u636e\u96c6\u5177\u6709\u53ef\u89e3\u91ca\u4fe1\u606f\uff0c\u4f5c\u8005\u5206\u522b\u518d\u65b0\u751f\u6210\u7684\u60c5\u611f\u6570\u636e\u96c6\u548c\u539f\u59cb\u6570\u636e\u96c6\u8fdb\u884c\u5b9e\u9a8c\u3002\u4f5c\u8005\u5c55\u793a\u4e86\u4e24\u5c42GNNs\u5728\u8fd9\u4e9b\u6570\u636e\u96c6\u4e0a\u7684\u9884\u6d4b\u7cbe\u5ea6\uff0c\u5305\u62ecGCNs\u3001GATs\u548cGINs\u3002\u8fd8\u5c55\u793a\u4e86\u4f7f\u7528\u539f\u59cb\u53e5\u5b50\u6570\u636e\u96c6\u7684\u9884\u8bad\u7ec3\u7684BERT[72]\u7684\u5fae\u8c03\u7cbe\u5ea6\u3002\u7ed3\u679c\u8868\u660e\uff0c\u4e0e\u539f\u59cb\u53e5\u5b50\u6570\u636e\u96c6\u76f8\u6bd4\uff0c\u4f5c\u8005\u6784\u5efa\u7684\u60c5\u611f\u56fe\u6570\u636e\u96c6\u53ef\u4ee5\u8fbe\u5230\u5177\u6709\u7ade\u4e89\u529b\u7684\u6027\u80fd<\/strong>\u8fd9\u4e9b\u6570\u636e\u96c6\u662f\u5b9e\u73b0\u56fe\u6a21\u578b\u89e3\u91ca\u7684\u5408\u7406\u9009\u62e9\u3002\u6839\u636e\u4e0d\u540c\u8bcd\u7684\u8bed\u4e49\u548c\u60c5\u611f\u6807\u7b7e\uff0c\u6211\u4eec\u53ef\u4ee5\u7814\u7a76\u53ef\u89e3\u91ca\u65b9\u6cd5\u662f\u5426\u80fd\u8bc6\u522b\u51fa\u5177\u6709\u5173\u952e\u610f\u4e49\u7684\u8bcd\u4ee5\u53ca\u4e0d\u540c\u8bcd\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/strong><\/span><\/section>\n

            5.1.3 Molecule data<\/span><\/h3>\n
            \u5206\u5b50\u6570\u636e\u96c6\u4e5f\u88ab\u5e7f\u6cdb\u7528\u4e8e\u89e3\u91ca\u4efb\u52a1\uff0c\u5982MUTAG[73]\u3001BBBP\u548cTox21[74]\u3002\u8fd9\u7c7b\u6570\u636e\u96c6\u4e2d\u7684\u6bcf\u4e2a\u56fe\u5bf9\u5e94\u4e00\u4e2a\u5206\u5b50\uff0c\u5176\u4e2d\u8282\u70b9\u4ee3\u8868\u539f\u5b50\uff0c\u8fb9\u662f\u5316\u5b66\u952e\u3002\u5206\u5b50\u56fe\u7684\u6807\u7b7e\u4e00\u822c\u7531\u5206\u5b50\u7684\u5316\u5b66\u529f\u80fd\u6216\u6027\u8d28\u51b3\u5b9a\u3002\u91c7\u7528\u8fd9\u6837\u7684\u6570\u636e\u96c6\u8fdb\u884c\u89e3\u91ca\u4efb\u52a1\u9700\u8981\u9886\u57df\u77e5\u8bc6\uff0c\u4f8b\u5982\u4ec0\u4e48\u5316\u5b66\u57fa\u56e2\u5bf9\u5176\u5b98\u80fd\u6027\u66f4\u5177\u6709\u9274\u522b\u6027\u3002\u4f8b\u5982\uff0c\u5728\u6570\u636e\u96c6MUTAG\u4e2d\uff0c\u4e0d\u540c\u7684\u56fe\u5f62\u662f\u6839\u636e\u5b83\u4eec\u5bf9\u7ec6\u83cc\u7684\u8bf1\u53d8\u4f5c\u7528\u6765\u6807\u6ce8\u7684\u3002\u4f8b\u5982\uff0c\u5df2\u77e5\u78b3\u73af\u548cNO2\u5316\u5b66\u57fa\u56e2\u53ef\u80fd\u5bfc\u81f4\u8bf1\u53d8\u6548\u5e94[73]\uff0c\u90a3\u4e48\u53ef\u4ee5\u7814\u7a76\u53ef\u89e3\u91ca\u65b9\u6cd5\u662f\u5426\u80fd\u8bc6\u522b\u51fa\u5bf9\u5e94\u7c7b\u522b\u7684 patterns\u3002<\/strong><\/span><\/section>\n
            \uff08\u5728\u4e0d\u540c\u7684\u9886\u57df\u4e2d\uff0c\u4e0d\u540c\u7684\u5c40\u90e8\u7ed3\u6784\u662f\u5177\u6709\u533a\u5206\u529b\u7684\uff0c\u53ef\u89e3\u91ca\u65b9\u6cd5\u662f\u5426\u80fd\u591f\u8bc6\u522b\u8fd9\u4e9b\u6a21\u5f0f\uff1f\uff09<\/span><\/section>\n

            5.2 Evaluation Metrics<\/span><\/h3>\n
            \u5373\u4f7f\u53ef\u89c6\u5316\u7684\u7ed3\u679c\u53ef\u4ee5\u8ba9\u4eba\u7406\u89e3\u89e3\u91ca\u6027\u65b9\u6cd5\u662f\u5426\u5408\u7406\uff0c\u4f46\u7531\u4e8e\u7f3a\u4e4f ground truths\uff0c\u8fd9\u79cd\u8bc4\u4f30\u5e76\u4e0d\u5b8c\u5168\u53ef\u4fe1\u3002\u4e3a\u4e86\u6bd4\u8f83\u4e0d\u540c\u7684\u89e3\u91ca\u6027\u65b9\u6cd5\uff0c\u6211\u4eec\u9700\u8981\u7814\u7a76\u6bcf\u4e2a\u8f93\u5165\u6837\u4f8b\u7684\u7ed3\u679c\uff0c\u8fd9\u5f88\u8017\u65f6\u3002\u56e0\u6b64\u8bc4\u4f30\u5ea6\u91cf\u5bf9\u4e8e\u7814\u7a76\u53ef\u89e3\u91ca\u65b9\u6cd5\u81f3\u5173\u91cd\u8981\u3002\u597d\u7684\u5ea6\u91cf\u65b9\u6cd5\u5e94\u8be5\u4ece\u6a21\u578b\u7684\u89d2\u5ea6\u6765\u8bc4\u4f30\u9884\u6d4b\u7ed3\u679c\uff0c\u6bd4\u5982\u89e3\u91ca\u662f\u5426\u5fe0\u5b9e\u4e8e\u6a21\u578b[75]\uff0c[76]\u3002\u4f5c\u8005\u5c06\u4ecb\u7ecd\u6700\u8fd1\u63d0\u51fa\u7684\u51e0\u79cd\u9488\u5bf9\u89e3\u91ca\u6027\u95ee\u9898\u7684\u8bc4\u4f30\u5ea6\u91cf\u65b9\u6cd5\u3002<\/span><\/section>\n

            5.2.1 Fidelity\/Infidelity<\/span><\/h3>\n
            \u9996\u5148\uff0c\u4ece\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u4e0a\u5206\u6790\u89e3\u91ca\u6027\u65b9\u6cd5\u7684\u6027\u80fd\uff0c\u89e3\u91ca\u5e94\u8be5\u5fe0\u4e8e\u6a21\u578b\uff0c\u89e3\u91ca\u65b9\u6cd5\u5e94\u8be5\u8bc6\u522b\u5bf9\u6a21\u578b\u91cd\u8981\u7684\u8f93\u5165\u7279\u5f81<\/strong>\u3002\u4e3a\u4e86\u8bc4\u4f30\u8fd9\u4e00\u70b9\uff0c\u6700\u8fd1\u63d0\u51fa\u4e86Fidelity[50]\u5ea6\u91cf\u65b9\u6cd5\u3002\u5173\u952e\u601d\u60f3\u5728\u4e8e\u5982\u679c\u89e3\u91ca\u6280\u672f\u6240\u8bc6\u522b\u7684\u91cd\u8981\u8f93\u5165\u7279\u5f81\uff08\u8282\u70b9\/\u8fb9\/\u8282\u70b9\u7279\u5f81\uff09\u5bf9\u6a21\u578b\u5177\u6709\u5224\u522b\u529b\uff0c\u90a3\u4e48\u5f53\u8fd9\u4e9b\u7279\u5f81\u88ab\u79fb\u9664\u65f6\uff0c\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u5e94\u8be5\u4f1a\u53d1\u751f\u663e\u8457\u53d8\u5316\u3002\u56e0\u6b64\uff0cFidelity\u88ab\u5b9a\u4e49\u4e3a\u539f\u59cb\u9884\u6d4b\u4e0e\u906e\u853d\u6389\u91cd\u8981\u8f93\u5165\u7279\u5f81\u540e\u7684\u65b0\u9884\u6d4b\u4e4b\u95f4\u7684\u7cbe\u5ea6\u4e4b\u5dee<\/strong>[50]\uff0c[77]\uff0c\u5373\u8861\u91cf\u4e24\u79cd\u9884\u6d4b\u7ed3\u679c\u7684\u5dee\u5f02\u6027\u3002<\/span><\/section>\n
            \u53ef\u89e3\u91ca\u65b9\u6cd5\u53ef\u4ee5\u770b\u4f5c\u662f\u4e00\u4e2a\u786c\u91cd\u8981\u6027\u6620\u5c04 <\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\uff0c\u5176\u4e2d\u5143\u7d20\u4e3a 0<\/span>\uff08\u8868\u793a\u7279\u5f81\u4e0d\u91cd\u8981\uff09<\/span>\u62161<\/span>\uff08\u8868\u793a\u7279\u5f81\u91cd\u8981\uff09<\/span>\u3002\u5bf9\u4e8e\u73b0\u6709\u65b9\u6cd5\uff0c\u4f8b\u5982ZORRO[51] \u548c Causal Screening[53] \u7b49\u65b9\u6cd5\uff0c\u751f\u6210\u7684\u89e3\u91ca\u662f\u79bb\u6563\u63a9\u7801\uff0c\u53ef\u4ee5\u76f4\u63a5\u4f5c\u4e3a\u91cd\u8981\u6027\u6620\u5c04 <\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u3002\u5bf9\u4e8e GNNExplainer[42] \u548c GraphLime[56] \u7b49\u65b9\u6cd5\uff0c\u91cd\u8981\u6027\u5206\u6570\u662f\u8fde\u7eed\u503c\uff0c\u90a3\u4e48\u53ef\u4ee5\u901a\u8fc7\u5f52\u4e00\u5316\u548c\u9608\u503c\u5316\u5f97\u5230\u91cd\u8981\u6027\u5730\u6620\u5c04<\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u3002\u6700\u540e\uff0c\u9884\u6d4b\u7cbe\u5ea6\u7684Fidelity\u5f97\u5206\u53ef\u4ee5\u8ba1\u7b97\u4e3a:<\/span><\/section>\n
            <\/p>\n
            <\/embed><\/section>\n

            <\/span><\/section>\n

            \u5176\u4e2d<\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u662f\u56fe<\/path><\/g><\/g><\/g><\/svg><\/span>\u7684\u539f\u59cb\u9884\u6d4b\uff0c<\/path><\/g><\/g><\/g><\/svg><\/span>\u662f\u56fe\u7684\u6570\u91cf\u3002<\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u8868\u793a\u53bb\u6389\u91cd\u8981\u8f93\u5165\u7279\u5f81\u7684\u8865\u5168\u63a9\u7801\uff0c<\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u662f\u5c06\u65b0\u56fe\u8f93\u5165\u8bad\u7ec3\u597d\u7684GNN <\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/svg><\/span>\u65f6\u7684\u9884\u6d4b\u503c\u3002\u6307\u793a\u51fd\u6570<\/path><\/g><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/svg><\/span>\u5982\u679c<\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u548c<\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u76f8\u7b49\u5219\u8fd4\u56de1\uff0c\u5426\u5219\u8fd4\u56de0\u3002\u6ce8\u610f\uff0c<\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u6307\u6807\u7814\u7a76\u7684\u662f\u9884\u6d4b\u7cbe\u5ea6\u7684\u53d8\u5316\u3002\u901a\u8fc7\u5bf9\u9884\u6d4b\u6982\u7387\u7684\u5173\u6ce8\uff0c\u6982\u7387\u7684Fidelity\u53ef\u4ee5\u5b9a\u4e49\u4e3a:<\/span><\/section>\n
            <\/p>\n
            <\/embed><\/section>\n

            <\/span><\/section>\n

            \u5176\u4e2d\uff0c<\/path><\/g><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u4ee3\u8868\u57fa\u4e8e\u4e92\u8865\u63a9\u7801<\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\uff0c\u4fdd\u7559<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u7684\u7279\u5f81\u5f97\u5230\u7684\u65b0\u56fe\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c<\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/g><\/path><\/path><\/path><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u76d1\u6d4b\u7684\u662f\u9884\u6d4b\u6982\u7387\u7684\u53d8\u5316\uff0c\u6bd4<\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u66f4\u654f\u611f\u3002\u5bf9\u4e8e\u8fd9\u4e24\u4e2a\u6307\u6807\u6765\u8bf4\uff0c\u6570\u503c\u8d8a\u9ad8\uff0c\u8bf4\u660e\u89e3\u91ca\u7ed3\u679c\u8d8a\u597d\uff0c\u8bc6\u522b\u51fa\u7684\u5224\u522b\u7279\u5f81\u8d8a\u591a<\/strong>\u3002<\/span><\/section>\n
            Fidelity\u5ea6\u91cf\u901a\u8fc7\u53bb\u9664\u91cd\u8981\u8282\u70b9\/\u8fb9\/\u8282\u70b9\u7279\u5f81\u6765\u7814\u7a76\u9884\u6d4b\u53d8\u5316<\/strong>\u3002\u76f8\u53cd\uff0cInfidelity\u5ea6\u91cf\u901a\u8fc7\u4fdd\u7559\u91cd\u8981\u7684\u8f93\u5165\u7279\u5f81\u548c\u53bb\u9664\u4e0d\u91cd\u8981\u7684\u7279\u5f81\u6765\u7814\u7a76\u9884\u6d4b\u53d8\u5316<\/strong>\u3002\u76f4\u89c2\u5730\u8bb2\uff0c\u91cd\u8981\u7279\u5f81\u5e94\u8be5\u5305\u542b\u5224\u522b\u4fe1\u606f\uff0c\u56e0\u6b64\u5373\u4f7f\u53bb\u6389\u4e0d\u91cd\u8981\u7684\u7279\u5f81\uff0c\u5b83\u4eec\u4e5f\u5e94\u8be5\u5bfc\u81f4\u4e0e\u539f\u59cb\u9884\u6d4b\u76f8\u4f3c\u7684\u9884\u6d4b\u3002\u4ece\u5f62\u5f0f\u4e0a\u770b\uff0c\u5ea6\u91cfInfidelity\u53ef\u4ee5\u8ba1\u7b97\u4e3a\uff1a<\/span><\/section>\n
            <\/p>\n
            <\/embed><\/section>\n

            <\/span><\/section>\n

            \u5176\u4e2d<\/path><\/g><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u662f\u6839\u636e\u6620\u5c04<\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u4fdd\u7559<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u7684\u91cd\u8981\u7279\u5f81\u65f6\u7684\u65b0\u56fe\uff0c<\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u662f\u65b0\u7684\u9884\u6d4b\u503c\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5bf9\u4e8e<\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/g><\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u548c<\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/path><\/g><\/path><\/path><\/path><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u6765\u8bf4\uff0c\u6570\u503c\u8d8a\u4f4e\uff0c\u8bf4\u660e\u53bb\u6389\u7684\u7279\u5f81\u91cd\u8981\u4fe1\u606f\u8d8a\u5c11\uff0c\u8fd9\u6837\u89e3\u91ca\u7ed3\u679c\u8d8a\u597d\u3002<\/strong><\/span><\/section>\n

            5.2.2 Sparsity<\/span><\/h3>\n
            \u4ece\u8f93\u5165\u56fe\u6570\u636e\u7684\u89d2\u5ea6\u6765\u5206\u6790\u89e3\u91ca\u6027\u65b9\u6cd5\u7684\u6027\u80fd\uff0c\u89e3\u91ca\u6027\u65b9\u6cd5\u5e94\u8be5\u662f\u7a00\u758f\u7684\uff0c\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u5e94\u8be5\u6355\u6349\u6700\u91cd\u8981\u7684\u8f93\u5165\u7279\u5f81\uff0c\u800c\u5ffd\u7565\u4e0d\u76f8\u5173\u7684\u7279\u5f81\uff0c\u53ef\u4ee5\u7528\u7a00\u758f\u5ea6\uff08Sparsity\uff09\u6307\u6807\u8861\u91cf\u8fd9\u6837\u4e2a\u7279\u6027\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u8861\u91cf\u7684\u662f\u88ab\u89e3\u91ca\u65b9\u6cd5\u9009\u62e9\u4e3a\u91cd\u8981\u7279\u5f81\u7684\u5206\u6570[50]\u3002\u5f62\u5f0f\u4e0a\uff0c\u7ed9\u5b9a\u56fe<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u548c\u5b83\u7684\u786c\u91cd\u8981\u6027\u6620\u5c04 <\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\uff0c\u7a00\u758f\u5ea6\u5ea6\u91cf\u53ef\u4ee5\u8ba1\u7b97\u4e3a\uff1a<\/span><\/section>\n
            <\/p>\n
            <\/embed><\/section>\n

            <\/span><\/section>\n

            \u5176\u4e2d<\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/svg><\/span>\u8868\u793a<\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u4e2d\u8bc6\u522b\u7684\u91cd\u8981\u8f93\u5165\u7279\u5f81\uff08\u8282\u70b9\/\u8fb9\/\u8282\u70b9\u7279\u5f81\uff09\u7684\u6570\u91cf\uff0c<\/path><\/g><\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/svg><\/span>\u8868\u793a\u539f\u59cb\u56fe <\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u4e2d\u7279\u5f81\u7684\u603b\u6570\u3002\u8bf7\u6ce8\u610f\uff0c\u6570\u503c\u8d8a\u9ad8\u8868\u793a\u89e3\u91ca\u65b9\u6cd5\u8d8a\u7a00\u758f\uff0c\u5373\u5f80\u5f80\u53ea\u6355\u6349\u6700\u91cd\u8981\u7684\u8f93\u5165\u4fe1\u606f\u3002<\/strong><\/span><\/section>\n

            5.2.3 Stability<\/span><\/h3>\n
            \u597d\u7684\u89e3\u91ca\u5e94\u8be5\u662f\u7a33\u5b9a\u7684\u3002\u5f53\u5bf9\u8f93\u5165\u65bd\u52a0\u5c0f\u7684\u53d8\u5316\u800c\u4e0d\u5f71\u54cd\u9884\u6d4b\u65f6\uff0c\u89e3\u91ca\u5e94\u8be5\u4fdd\u6301\u76f8\u4f3c\u3002\u6700\u8fd1\u63d0\u51fa\u7684\u7a33\u5b9a\u6027\u5ea6\u91cf\u6807\u51c6\u6765\u8861\u91cf\u4e00\u4e2a\u89e3\u91ca\u65b9\u6cd5\u662f\u5426\u7a33\u5b9a[78]\u3002\u7ed9\u5b9a\u4e00\u4e2a\u8f93\u5165\u56fe<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\uff0c\u5b83\u7684\u89e3\u91ca<\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u88ab\u8ba4\u4e3a\u662f\u771f\u5b9e\u6807\u7b7e\u3002\u7136\u540e\u5bf9\u8f93\u5165\u56fe<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u8fdb\u884c\u5c0f\u7684\u6539\u53d8\uff0c\u6bd4\u5982\u9644\u52a0\u65b0\u7684\u8282\u70b9\/\u8fb9\uff0c\u5f97\u5230\u4e00\u4e2a\u65b0\u7684\u56fe<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u548c<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u9700\u8981\u6709\u76f8\u540c\u7684\u9884\u6d4b\u3002\u7136\u540e\u5f97\u5230<\/path><\/g><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span>\u7684\u89e3\u91ca\uff0c\u8868\u793a\u4e3a<\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg><\/span> \u3002\u901a\u8fc7\u6bd4\u8f83<\/strong><\/span><\/path><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg>\u548c<\/strong><\/span><\/path><\/g><\/path><\/g><\/g><\/g><\/path><\/g><\/g><\/g><\/g><\/g><\/svg>\u4e4b\u95f4\u7684\u5dee\u5f02<\/strong>\uff0c\u6211\u4eec\u53ef\u4ee5\u8ba1\u7b97\u51fa\u7a33\u5b9a\u6027\u5f97\u5206\u3002\u8bf7\u6ce8\u610f\uff0c\u6570\u503c\u8d8a\u4f4e\u8868\u793a\u89e3\u91ca\u6280\u672f\u8d8a\u7a33\u5b9a\uff0c\u5bf9\u566a\u58f0\u4fe1\u606f\u7684\u9c81\u68d2\u6027\u8d8a\u5f3a\u3002<\/strong><\/span><\/section>\n

            5.2.4 Accuracy<\/span><\/h3>\n
            \u9488\u5bf9\u5408\u6210\u6570\u636e\u96c6\u63d0\u51fa\u4e86\u7cbe\u5ea6\u5ea6\u91cf\u65b9\u6cd5[42]\u3001[78]\u3002\u5728\u5408\u6210\u6570\u636e\u96c6\u4e2d\uff0c\u5373\u4f7f\u4e0d\u77e5\u9053GNN\u662f\u5426\u6309\u7167\u6211\u4eec\u9884\u671f\u7684\u65b9\u5f0f\u8fdb\u884c\u9884\u6d4b\uff0c\u4f46\u6784\u5efa\u8fd9\u4e9b\u6570\u636e\u96c6\u7684\u89c4\u5219\uff0c\u5982 graph motifs\uff0c\u53ef\u4ee5\u4f5c\u4e3a ground truths \u7684\u5408\u7406\u8fd1\u4f3c\u3002\u7136\u540e\u5bf9\u4e8e\u4efb\u4f55\u8f93\u5165\u56fe\uff0c\u6211\u4eec\u90fd\u53ef\u4ee5\u5c06\u5176\u89e3\u91ca\u4e0e\u8fd9\u6837\u7684 ground truths\u8fdb\u884c\u6bd4\u8f83\u3002\u4f8b\u5982\uff0c\u5728\u7814\u7a76\u91cd\u8981\u8fb9\u7684\u65f6\u5019\uff0c\u53ef\u4ee5\u7814\u7a76\u89e3\u91ca\u4e2d\u7684\u91cd\u8981\u8fb9\u4e0e ground truths \u7684\u8fb9\u7684\u5339\u914d\u7387\u3002\u8fd9\u79cd\u6bd4\u8f83\u7684\u5e38\u7528\u6307\u6807\u5305\u62ec\u4e00\u822c\u7cbe\u5ea6\u3001F1\u5f97\u5206\u3001ROC-AUC\u5f97\u5206\u3002\u5339\u914d\u7387\u6570\u503c\u8d8a\u9ad8\uff0c\u8bf4\u660e\u89e3\u91ca\u7ed3\u679c\u8d8a\u63a5\u8fd1\u4e8e ground truths\uff0c\u8ba4\u4e3a\u662f\u8f83\u597d\u7684\u89e3\u91ca\u65b9\u6cd5\u3002<\/strong><\/span><\/section>\n
            <\/p>\n
            \n

             <\/span><\/p>\n<\/section>\n

            <\/span><\/section>\n

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

            <\/section>\n
            \n
            <\/svg><\/section>\n<\/section>\n<\/section>\n
            \n

            6. Conclusion<\/strong><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

            <\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n
            \u56fe\u795e\u7ecf\u7f51\u7edc\u8fd1\u6765\u88ab\u5e7f\u6cdb\u7814\u7a76\uff0c\u4f46\u5bf9\u56fe\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u7684\u63a2\u8ba8\u8fd8\u6bd4\u8f83\u5c11\u3002\u4e3a\u4e86\u7814\u7a76\u8fd9\u4e9b\u9ed1\u7bb1\u7684\u6f5c\u5728\u673a\u5236\uff0c\u4eba\u4eec\u63d0\u51fa\u4e86\u51e0\u79cd\u89e3\u91ca\u56fe\u6a21\u578b\u7684\u65b9\u6cd5\uff0c\u5305\u62ecXGNN\u3001GNNExplainer\u7b49\u3002\u8fd9\u4e9b\u65b9\u6cd5\u4ece\u4e0d\u540c\u7684\u89d2\u5ea6\u548c\u52a8\u673a\u6765\u89e3\u91ca\u56fe\u6a21\u578b\uff0c\u4f46\u662f\u7f3a\u4e4f\u5bf9\u8fd9\u4e9b\u65b9\u6cd5\u7684\u5168\u9762\u7814\u7a76\u548c\u5206\u6790\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u4f5c\u8005\u5bf9\u8fd9\u4e9b\u65b9\u6cd5\u8fdb\u884c\u4e86\u7cfb\u7edf\u5168\u9762\u7684\u8c03\u7814\u3002\u9996\u5148\u5bf9\u73b0\u6709\u7684GNN\u89e3\u91ca\u65b9\u6cd5\u8fdb\u884c\u4e86\u7cfb\u7edf\u7684\u5206\u7c7b\uff0c\u5e76\u4ecb\u7ecd\u4e86\u6bcf\u4e00\u7c7b\u89e3\u91ca\u65b9\u6cd5\u80cc\u540e\u7684\u5173\u952e\u601d\u60f3\u3002\u7136\u540e\u8be6\u7ec6\u8ba8\u8bba\u4e86\u6bcf\u4e00\u79cd\u89e3\u91ca\u65b9\u6cd5\uff0c\u5305\u62ec\u65b9\u6cd5\u3001\u5185\u6db5\u3001\u4f18\u52bf\u548c\u7f3a\u70b9\uff0c\u8fd8\u5bf9\u4e0d\u540c\u7684\u89e3\u91ca\u65b9\u6cd5\u8fdb\u884c\u4e86\u7efc\u5408\u5206\u6790\u3002\u5e76\u4e14\u4ecb\u7ecd\u548c\u5206\u6790\u4e86\u5e38\u7528\u7684\u6570\u636e\u96c6\u548cGNN\u89e3\u91ca\u65b9\u6cd5\u7684\u8bc4\u4ef7\u6307\u6807\u3002\u6700\u540e\u4ece\u6587\u672c\u6570\u636e\u51fa\u53d1\uff0c\u5efa\u7acb\u4e86\u4e09\u4e2a\u56fe\u5f62\u6570\u636e\u96c6\uff0c\u8fd9\u4e9b\u6570\u636e\u96c6\u662f\u4eba\u7c7b\u53ef\u4ee5\u7406\u89e3\u7684\uff0c\u53ef\u4ee5\u76f4\u63a5\u7528\u4e8eGNN\u89e3\u91ca\u4efb\u52a1\u3002<\/span><\/section>\n
            \n
            \n
            \n
            \n
            \n

            \u53c2\u8003\u6587\u732e<\/strong><\/span><\/p>\n<\/section>\n<\/section>\n<\/section>\n

            \n
            \n
            \n
            \n

            [10] K. Simonyan, A. Vedaldi, and A. Zisserman, \u201cDeep inside convolutional networks: Visualising image classification models and saliency maps,\u201d arXiv preprint arXiv:1312.6034, 2013.<\/span><\/p>\n

            [11] D. Smilkov, N. Thorat, B. Kim, F. Viegas, and M. Wattenberg, \u00b4 \u201cSmoothgrad: removing noise by adding noise,\u201d arXiv preprint arXiv:1706.03825, 2017.<\/span><\/p>\n

            [12] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, \u201cLearning deep features for discriminative localization,\u201d in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 2921\u20132929.<\/span><\/p>\n

            [13] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, \u201cGrad-cam: Visual explanations from deep networks via gradient-based localization,\u201d in 2017 IEEE International Conference on Computer Vision (ICCV). IEEE, 2017, pp. 618\u2013626.<\/span><\/p>\n

            [14] P. Dabkowski and Y. Gal, \u201cReal time image saliency for black box classifiers,\u201d in Advances in Neural Information Processing Systems, 2017, pp. 6967\u20136976.<\/span><\/p>\n

            [15] H. Yuan, L. Cai, X. Hu, J. Wang, and S. Ji, \u201cInterpreting image classifiers by generating discrete masks,\u201d IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020.<\/span><\/p>\n

            [17] C. Olah, A. Satyanarayan, I. Johnson, S. Carter, L. Schubert, K. Ye, and A. Mordvintsev, \u201cThe building blocks of interpretability,\u201d Distill, 2018, https:\/\/distill.pub\/2018\/building-blocks.<\/span><\/p>\n

            [18] F. Yang, S. K. Pentyala, S. Mohseni, M. Du, H. Yuan, R. Linder, E. D. Ragan, S. Ji, and X. Hu, \u201cXfake: explainable fake news detector with visualizations,\u201d in The World Wide Web Conference, 2019, pp. 3600\u20133604.<\/span><\/p>\n

            [19] M. Du, N. Liu, Q. Song, and X. Hu, \u201cTowards explanation of dnnbased prediction with guided feature inversion,\u201d in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2018, pp. 1358\u20131367.<\/span><\/p>\n

            [22] H. Yuan, Y. Chen, X. Hu, and S. Ji, \u201cInterpreting deep models for text analysis via optimization and regularization methods,\u201d in AAAI-19: Thirty-Third AAAI Conference on Artificial Intelligence. Association for the Advancement of Artificial Intelligence, 2019.<\/span><\/p>\n

            [23] M. Du, N. Liu, and X. Hu, \u201cTechniques for interpretable machine learning,\u201d Communications of the ACM, vol. 63, no. 1, pp. 68\u201377, 2019.<\/span><\/p>\n

            [24] A. Rai, \u201cExplainable ai: From black box to glass box,\u201d Journal of the Academy of Marketing Science, vol. 48, no. 1, pp. 137\u2013141, 2020.<\/span><\/p>\n

            [25] F. K. Dosilovi \u02c7 c, M. Br \u00b4 ci\u02c7 c, and N. Hlupi \u00b4 c, \u201cExplainable artificial \u00b4 intelligence: A survey,\u201d in 2018 41st International convention on information and communication technology, electronics and microelectronics (MIPRO). IEEE, 2018, pp. 0210\u20130215. [26] C. Molnar, Interpretable Machine Learning, 2019, https:\/\/ christophm.github.io\/interpretable-ml-book\/.<\/span><\/p>\n

            [41] H. Yuan, J. Tang, X. Hu, and S. Ji, \u201cXGNN: Towards model-level explanations of graph neural networks,\u201d ser. KDD \u201920. New York, NY, USA: Association for Computing Machinery, 2020, p. 430\u2013438. [Online]. Available: https:\/\/doi.org\/10.1145\/3394486.3403085<\/span><\/p>\n

            [42] Z. Ying, D. Bourgeois, J. You, M. Zitnik, and J. Leskovec, \u201cGnnexplainer: Generating explanations for graph neural networks,\u201d in Advances in neural information processing systems, 2019, pp. 9244\u2013 9255.<\/span><\/p>\n

            [43] D. Luo, W. Cheng, D. Xu, W. Yu, B. Zong, H. Chen, and X. Zhang, \u201cParameterized explainer for graph neural network,\u201d in Advances in neural information processing systems, 2020.<\/span><\/p>\n

            [44] C. Rudin, \u201cStop explaining black box machine learning models for high stakes decisions and use interpretable models instead,\u201d Nature Machine Intelligence, vol. 1, no. 5, pp. 206\u2013215, 2019.<\/span><\/p>\n

            [45] J. Chen, L. Song, M. J. Wainwright, and M. I. Jordan, \u201cLearning to explain: An information-theoretic perspective on model interpretation,\u201d in International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 80. PMLR, 2018, pp. 882\u2013891.<\/span><\/p>\n

            [46] U. Alon, \u201cNetwork motifs: theory and experimental approaches,\u201d Nature Reviews Genetics, vol. 8, no. 6, pp. 450\u2013461, 2007.<\/span><\/p>\n

            [47] R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon, \u201cNetwork motifs: simple building blocks of complex networks,\u201d Science, vol. 298, no. 5594, pp. 824\u2013827, 2002.<\/span><\/p>\n

            [48] U. Alon, An introduction to systems biology: design principles of biological circuits. CRC press, 2019.<\/span><\/p>\n

            [49] F. Baldassarre and H. Azizpour, \u201cExplainability techniques for graph convolutional networks,\u201d in International Conference on Machine Learning (ICML) Workshops, 2019 Workshop on Learning and Reasoning with Graph-Structured Representations, 2019.<\/span><\/p>\n

            [50] P. E. Pope, S. Kolouri, M. Rostami, C. E. Martin, and H. Hoffmann, \u201cExplainability methods for graph convolutional neural networks,\u201d in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 10 772\u201310 781.<\/span><\/p>\n

            [51] Anonymous, \u201cHard masking for explaining graph neural networks,\u201d in Submitted to International Conference on Learning Representations, 2021, under review. [Online]. Available: https: \/\/openreview.net\/forum?id=uDN8pRAdsoC<\/span><\/p>\n

            [52] M. S. Schlichtkrull, N. De Cao, and I. Titov, \u201cInterpreting graph neural networks for nlp with differentiable edge masking,\u201d arXiv preprint arXiv:2010.00577, 2020.<\/span><\/p>\n

            [53] Anonymous, \u201cCausal screening to interpret graph neural networks,\u201d in Submitted to International Conference on Learning Representations, 2021, under review. [Online]. Available: https: \/\/openreview.net\/forum?id=nzKv5vxZfge<\/span><\/p>\n

            [54] R. Schwarzenberg, M. Hubner, D. Harbecke, C. Alt, and L. Hennig, \u00a8 \u201cLayerwise relevance visualization in convolutional text graph classifiers,\u201d arXiv preprint arXiv:1909.10911, 2019.<\/span><\/p>\n

            [55] T. Schnake, O. Eberle, J. Lederer, S. Nakajima, K. T. Schutt, K.-R. \u00a8 Muller, and G. Montavon, \u201cHigher-order explanations of graph \u00a8 neural networks via relevant walks,\u201d 2020.<\/span><\/p>\n

            [56] Q. Huang, M. Yamada, Y. Tian, D. Singh, D. Yin, and Y. Chang, \u201cGraphlime: Local interpretable model explanations for graph neural networks,\u201d arXiv preprint arXiv:2001.06216, 2020.<\/span><\/p>\n

            [57] Y. Zhang, D. Defazio, and A. Ramesh, \u201cRelex: A model-agnostic relational model explainer,\u201d arXiv preprint arXiv:2006.00305, 2020.<\/span><\/p>\n

            [58] M. N. Vu and M. T. Thai, \u201cPgm-explainer: Probabilistic graphical model explanations for graph neural networks,\u201d in Advances in neural information processing systems, 2020.<\/span><\/p>\n

            [59] A. Shrikumar, P. Greenside, and A. Kundaje, \u201cLearning important features through propagating activation differences,\u201d in International Conference on Machine Learning, 2017, pp. 3145\u20133153.<\/span><\/p>\n

            [60] J. Chen, L. Song, M. Wainwright, and M. Jordan, \u201cLearning to explain: An information-theoretic perspective on model interpretation,\u201d in Proceedings of the 35th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, J. Dy and A. Krause, Eds., vol. 80. Stockholmsmassan, Stockholm \u00a8 Sweden: PMLR, 10\u201315 Jul 2018, pp. 883\u2013892. [Online]. Available: http:\/\/proceedings.mlr.press\/v80\/chen18j.html<\/span><\/p>\n

            [61] R. S. Sutton, D. McAllester, S. Singh, and Y. Mansour, \u201cPolicy gradient methods for reinforcement learning with function approximation,\u201d Advances in neural information processing systems, vol. 12, pp. 1057\u20131063, 1999.<\/span><\/p>\n

            [62] E. Jang, S. Gu, and B. Poole, \u201cCategorical reparameterization with gumbel-softmax,\u201d in International Conference on Learning Representations, 2016.<\/span><\/p>\n

            [63] C. Louizos, M. Welling, and D. P. Kingma, \u201cLearning sparse neural networks through l 0 regularization,\u201d arXiv preprint arXiv:1712.01312, 2017.<\/span><\/p>\n

            [64] M. T. Ribeiro, S. Singh, and C. Guestrin, \u201c\u201d why should i trust you?\u201d explaining the predictions of any classifier,\u201d in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp. 1135\u20131144.<\/span><\/p>\n

            [65] M. Yamada, W. Jitkrittum, L. Sigal, E. P. Xing, and M. Sugiyama, \u201cHigh-dimensional feature selection by feature-wise kernelized lasso,\u201d Neural computation, vol. 26, no. 1, pp. 185\u2013207, 2014.<\/span><\/p>\n

            [66] D. Margaritis and S. Thrun, \u201cBayesian network induction via local neighborhoods,\u201d Advances in neural information processing systems, vol. 12, pp. 505\u2013511, 1999.<\/span><\/p>\n

            [67] S. Bach, A. Binder, G. Montavon, F. Klauschen, K. Muller, and \u00a8 W. Samek, \u201cOn pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation,\u201d PLoS ONE, vol. 10, no. 7, p. e0130140, 2015.<\/span><\/p>\n

            [68] R. Albert and A.-L. Barabasi, \u201cStatistical mechanics of complex \u00b4 networks,\u201d Reviews of modern physics, vol. 74, no. 1, p. 47, 2002.<\/span><\/p>\n

            [69] R. Socher, A. Perelygin, J. Wu, J. Chuang, C. D. Manning, A. Y. Ng, and C. Potts, \u201cRecursive deep models for semantic compositionality over a sentiment treebank,\u201d in Proceedings of the 2013 conference on empirical methods in natural language processing, 2013, pp. 1631\u20131642.<\/span><\/p>\n

            [70] L. Dong, F. Wei, C. Tan, D. Tang, M. Zhou, and K. Xu, \u201cAdaptive recursive neural network for target-dependent twitter sentiment classification,\u201d in Proceedings of the 52nd annual meeting of the association for computational linguistics (volume 2: Short papers), 2014, pp. 49\u201354.<\/span><\/p>\n

            [71] M. Gardner, J. Grus, M. Neumann, O. Tafjord, P. Dasigi, N. Liu, M. Peters, M. Schmitz, and L. Zettlemoyer, \u201cAllennlp: A deep semantic natural language processing platform,\u201d arXiv preprint arXiv:1803.07640, 2018.<\/span><\/p>\n

            [72] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, \u201cBert: Pretraining of deep bidirectional transformers for language understanding,\u201d arXiv preprint arXiv:1810.04805, 2018.<\/span><\/p>\n

            [73] A. K. Debnath, R. L. Lopez de Compadre, G. Debnath, A. J. Shusterman, and C. Hansch, \u201cStructure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. correlation with molecular orbital energies and hydrophobicity,\u201d Journal of medicinal chemistry, vol. 34, no. 2, pp. 786\u2013797, 1991.<\/span><\/p>\n

            [74] Z. Wu, B. Ramsundar, E. N. Feinberg, J. Gomes, C. Geniesse, A. S. Pappu, K. Leswing, and V. Pande, \u201cMoleculenet: a benchmark for molecular machine learning,\u201d Chemical science, vol. 9, no. 2, pp. 513\u2013530, 2018.<\/span><\/p>\n

            [75] A. Jacovi and Y. Goldberg, \u201cTowards faithfully interpretable nlp systems: How should we define and evaluate faithfulness?\u201d arXiv preprint arXiv:2004.03685, 2020.<\/span><\/p>\n

            [76] S. Wiegreffe and Y. Pinter, \u201cAttention is not not explanation,\u201d arXiv preprint arXiv:1908.04626, 2019.<\/span><\/p>\n

            [77] S. Hooker, D. Erhan, P.-J. Kindermans, and B. Kim, \u201cA benchmark for interpretability methods in deep neural networks,\u201d in Advances in Neural Information Processing Systems, 2019, pp. 9737\u20139748.<\/span><\/p>\n

            [78] B. Sanchez-Lengeling, J. Wei, B. Lee, E. Reif, P. Wang, W. W. Qian, K. McCloskey, L. Colwell, and A. Wiltschko, \u201cEvaluating attribution for graph neural networks,\u201d Advances in Neural Information Processing Systems, vol. 33, 2020.<\/span><\/p>\n


            <\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

            \n
            \n

            \uff08\u53c2\u8003\u6587\u732e\u53ef\u4e0a\u4e0b\u6ed1\u52a8<\/strong>\u67e5\u770b\uff09<\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

            \n


            <\/p>\n<\/section>\n


            <\/section>\n

             <\/span><\/p>\n

            \n
            \n
            \n
            \n
            \n

            \u56e0\u679c\u79d1\u5b66\u7b2c\u4e8c\u5b63\u8bfb\u4e66\u4f1a\u62a5\u540d\u4e2d<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


            <\/section>\n<\/h2>\n

            \u56e0\u679c\u63a8\u65ad\u4e0e\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u7ed3\u5408\u5df2\u7ecf\u5438\u5f15\u4e86\u8d8a\u6765\u8d8a\u591a\u6765\u81ea\u5b66\u754c\u4e1a\u754c\u7684\u5173\u6ce8\uff0c\u4e3a\u6df1\u5165\u63a2\u8ba8\u3001\u666e\u53ca\u63a8\u5e7f\u56e0\u679c\u79d1\u5b66\u8bae\u9898\uff0c\u5e2e\u52a9\u5176\u4ed6\u9886\u57df\u5982\u7ecf\u6d4e\u5b66\u53ef\u4ee5\u5229\u7528\u673a\u5668\u5b66\u4e60\u7684\u8fdb\u6b65\u6765\u5e2e\u52a9\u56de\u7b54\u4ed6\u4eec\u81ea\u5df1\u7684\u56e0\u679c\u95ee\u9898<\/strong>\uff0c\u96c6\u667a\u4ff1\u4e50\u90e8\u643a\u624b\u667a\u6e90\u793e\u533a\u5c06\u4e3e\u529e\u7b2c\u4e8c\u5b63\u300c<\/span>\u56e0\u679c\u79d1\u5b66\u4e0eCausalAI\u8bfb\u4e66\u4f1a<\/span><\/a>\u300d\u3002\u672c\u671f\u8bfb\u4e66\u4f1a\u7740\u529b\u4e8e\u5b9e\u64cd\u6027\u3001\u57fa\u7840\u6027\uff0c\u5c06\u5e26\u9886\u5927\u5bb6\u7cbe\u8bfb\u56e0\u679c\u79d1\u5b66\u65b9\u5411\u4e24\u672c\u975e\u5e38\u53d7\u5e7f\u6cdb\u8ba4\u53ef\u7684\u5165\u95e8\u6559\u6750\u3002<\/span><\/p>\n


            <\/section>\n
            1. Pearl, Judea, Madelyn Glymour, and Nicholas P. Jewell. Causal inference in statistics: A primer. John Wiley & Sons, 2016.<\/span>\uff08\u672c\u4e66\u4e2d\u8bd1\u7248\u300a\u7edf\u8ba1\u56e0\u679c\u63a8\u7406\u5165\u95e8\uff08\u7ffb\u8bd1\u7248\uff09\u300b\u5df2\u7531\u9ad8\u7b49\u6559\u80b2\u51fa\u7248\u793e\u51fa\u7248\uff09<\/span><\/section>\n

            <\/section>\n
            2. Peters, Jonas, Dominik Janzing, and Bernhard Sch\u00f6lkopf. Elements of causal inference: foundations and learning algorithms. The MIT Press, 2017.<\/span><\/section>\n

            <\/section>\n
            \u8bfb\u4e66\u4f1a\u6bcf\u5468\u5c06\u8fdb\u884c\u76f4\u64ad\u8ba8\u8bba\uff0c\u8fdb\u884c\u95ee\u9898\u4ea4\u6d41\u3001\u91cd\u70b9\u6982\u5ff5\u5206\u4eab\u3001\u9605\u8bfb\u6982\u89c8\u548c\u7f16\u7a0b\u5b9e\u8df5\u5185\u5bb9\u5206\u6790\u3002\u975e\u5e38\u9002\u5408\u6709\u673a\u5668\u5b66\u4e60\u80cc\u666f\uff0c\u5e0c\u671b\u6df1\u5165\u5b66\u4e60\u56e0\u679c\u79d1\u5b66\u57fa\u7840\u77e5\u8bc6\u548c\u91cd\u8981\u6a21\u578b\u65b9\u6cd5\uff0c\u5bfb\u6c42\u89e3\u51b3\u76f8\u5173\u7814\u7a76\u95ee\u9898\u7684\u670b\u53cb\u53c2\u52a0\u3002<\/span><\/section>\n

            <\/section>\n
            \u76ee\u524d\u56e0\u679c\u79d1\u5b66\u8bfb\u4e66\u4f1a\u7cfb\u5217\uff0c\u5df2\u7ecf\u6709\u63a5\u8fd1400\u591a\u4f4d\u7684\u6d77\u5185\u5916\u9ad8\u6821\u79d1\u7814\u9662\u6240\u7684\u4e00\u7ebf\u79d1\u7814\u5de5\u4f5c\u8005\u4ee5\u53ca\u4e92\u8054\u7f51\u4e00\u7ebf\u4ece\u4e1a\u4eba\u5458\u53c2\u4e0e\uff0c\u5438\u5f15\u4e86\u56fd\u5185\u548c\u56fd\u9645\u4e0a\u5927\u90e8\u5206\u7684\u56e0\u679c\u79d1\u5b66\u9886\u57df\u7684\u4e13\u4e1a\u79d1\u7814\u4eba\u5458\uff0c\u5982\u679c\u4f60\u4e5f\u5bf9\u8fd9\u4e2a\u4e3b\u9898\u611f\u5174\u8da3\uff0c\u60f3\u8981\u6df1\u5ea6\u5730\u53c2\u4e0e\uff0c\u5c31\u5feb\u52a0\u5165\u6211\u4eec\u5427\uff01<\/span><\/section>\n

            <\/section>\n
            \u8be6\u60c5\u8bf7\u70b9\u51fb\uff1a<\/span><\/span><\/strong><\/section>\n
            \u8fde\u63a5\u7edf\u8ba1\u5b66\u3001\u673a\u5668\u5b66\u4e60\u4e0e\u81ea\u52a8\u63a8\u7406\u7684\u65b0\u5174\u4ea4\u53c9\u9886\u57df\u2014\u2014\u56e0\u679c\u79d1\u5b66\u8bfb\u4e66\u4f1a\u518d\u8d77\u822a<\/a><\/span><\/span><\/strong><\/section>\n

            <\/section>\n


            <\/p>\n

            \n

            \u63a8\u8350\u9605\u8bfb<\/span><\/strong><\/strong>
            <\/p>\n