{"id":32493,"date":"2022-02-17T22:54:24","date_gmt":"2022-02-17T14:54:24","guid":{"rendered":"https:\/\/swarma.org\/?p=32493"},"modified":"2022-02-17T22:54:24","modified_gmt":"2022-02-17T14:54:24","slug":"2022%e5%b9%b4%ef%bc%8c%e5%9b%be%e6%9c%ba%e5%99%a8%e5%ad%a6%e4%b9%a0graph-ml%e5%8f%91%e5%b1%95%e5%88%b0%e5%93%aa%e4%ba%86%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=32493","title":{"rendered":"2022\u5e74\uff0c\u56fe\u673a\u5668\u5b66\u4e60Graph ML\u53d1\u5c55\u5230\u54ea\u4e86\uff1f"},"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
\u5bf92021\u5e74 Graph ML \u8fdb\u5c55\u8fdb\u884c\u7ed3\u6784\u5316\u5206\u6790\uff0c\u91cd\u70b9\u4ecb\u7ecd\u70ed\u95e8\u8d8b\u52bf\u548c\u91cd\u5927\u8fdb\u6b65\u3002
<\/strong><\/section>\n

<\/strong><\/section>\n
\u7814\u7a76\u9886\u57df\uff1a\u56fe\u795e\u7ecf\u7f51\u7edc\uff0c\u673a\u5668\u5b66\u4e60
<\/strong><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

Michael Galkin<\/span>| \u4f5c\u8005<\/strong><\/strong><\/span><\/p>\n<\/section>\n

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


<\/section>\n


<\/h1>\n
\n
\n
\n

\u8bba\u6587\u9898\u76ee\uff1aGraph ML in 2022: Where are we now?<\/p>\n

\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/mgalkin.medium.com\/graph-ml-in-2022-where-are-we-now-f7f8242599e0<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n


<\/p>\n

\u5bf9\u4e8e Graph ML \u6765\u8bf4\u5df2\u7ecf\u8fc7\u53bb\u4e86\u4e00\u5e74\uff1a\u6210\u5343\u4e0a\u4e07\u7684\u8bba\u6587\uff0c\u65e0\u6570\u7684\u4f1a\u8bae\u548c\u7814\u8ba8\u4f1a\uff0c\u5982\u4f55\u8d76\u4e0a\u8fd9\u4e48\u591a\u5f88\u9177\u7684\u8fdb\u5c55\uff1f<\/span>
<\/section>\n

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


<\/p>\n

\u672c\u6587\u5bf9 Graph ML \u8fdb\u884c\u7ed3\u6784\u5316\u5206\u6790\uff0c\u91cd\u70b9\u4ecb\u7ecd\u70ed\u95e8\u8d8b\u52bf\u548c\u91cd\u5927\u8fdb\u6b65\u3002<\/span><\/p>\n


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


<\/h2>\n


<\/p>\n

\n
\n
\n
\n
\n

<\/section>\n

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

1.<\/strong><\/strong><\/span>Graph Transformers<\/span><\/strong><\/p>\n

+ Positional <\/span><\/strong>Features<\/span><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

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


<\/p>\n

\u867d\u7136\u56fe\u795e\u7ecf\u7f51\u7edc<\/span>\uff08Graph Neural Networks\uff0cGNN\uff09<\/span>\u901a\u5e38\u5728\u7a00\u758f\u56fe\u4e0a\u8fd0\u884c\uff0c\u4f46 Graph Transformers<\/span> (GT) <\/span>\u5728\u5168\u8fde\u63a5\u56fe\u4e0a\u8fd0\u884c\uff0c\u5176\u4e2d\u6bcf\u4e2a\u8282\u70b9\u90fd\u8fde\u63a5\u5230\u56fe\u4e2d\u7684\u6bcf\u4e2a\u5176\u4ed6\u8282\u70b9\u3002\u4e00\u65b9\u9762\uff0c\u8fd9\u5e26\u6765\u4e86\u8282\u70b9\u6570 N \u7684 O<\/span>(N\u00b2) <\/span>\u590d\u6742\u5ea6\u3002\u53e6\u4e00\u65b9\u9762\uff0cGT \u4e0d\u4f1a\u906d\u53d7\u8fc7\u5ea6\u5e73\u6ed1\uff0c\u8fd9\u662f\u957f\u8ddd\u79bb\u6d88\u606f\u4f20\u9012\u7684\u5e38\u89c1\u95ee\u9898\u3002<\/span>(\u4e3a\u4ec0\u4e48\u4e0d\u4f1a\u8fc7\u5e73\u6ed1\uff1f)<\/span><\/p>\n


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

\u5168\u8fde\u63a5\u56fe\u610f\u5473\u7740\u6211\u4eec\u6709\u6765\u81ea\u539f\u59cb\u56fe\u7684\u201c\u771f\u201d\u8fb9\u548c\u4ece\u5168\u8fde\u63a5\u53d8\u6362\u6dfb\u52a0\u7684\u201c\u5047\u201d\u8fb9\uff0c\u6211\u4eec\u60f3\u533a\u5206\u5b83\u4eec\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u6211\u4eec\u9700\u8981\u4e00\u79cd\u65b9\u6cd5\u6765\u4e3a\u8282\u70b9\u6ce8\u5165\u4e00\u4e9b\u4f4d\u7f6e\u7279\u5f81\uff0c\u5426\u5219 GT \u4f1a\u843d\u540e\u4e8e GNN<\/span>\uff08\u5982Dwivedi \u548c Bresson\u76842020 \u5e74\u8bba\u6587\u6240\u793a\uff09<\/span>\u3002\u4eca\u5e74\u6700\u5f15\u4eba\u6ce8\u76ee\u7684\u4e24\u4e2aGraph Transformers \u53ef\u80fd\u662f SAN<\/span>\uff08Spectral Attention Nets\uff09<\/span>\u548c Graphormer\u3002<\/span><\/p>\n


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

    \n
  • \n

    SAN\u91c7\u7528\u7684top-k\u7684\u62c9\u666e\u62c9\u65af\u7279\u5f81\u503c\u548c\u7279\u5f81\u5411\u91cf\uff0c\u5176\u53ef\u4ee5\u5355\u72ec\u533a\u5206\u75311-WL\u6d4b\u8bd5\u8003\u8651\u540c\u6784\u7684\u56fe\u3002SAN \u5c06\u5149\u8c31\u7279\u5f81\u4e0e\u8f93\u5165\u8282\u70b9\u7279\u5f81\u8fde\u63a5\u8d77\u6765\uff0c\u5728\u8bb8\u591a\u5206\u5b50\u4efb\u52a1\u4e0a\u4f18\u4e8e\u7a00\u758f GNN\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n


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


    <\/p>\n

      \n
    • \n
      Graphormer\u91c7\u7528\u4e0d\u540c\u7684\u65b9\u6cd5\u4e3b\u8981\u5728\u7a7a\u57df\u6765\u641e\u3002\u9996\u5148\uff0c\u8282\u70b9\u7279\u5f81\u8003\u8651\u4e86\u4e2d\u5fc3\u6027\u7f16\u7801\u2014\u2014\u53ef\u5b66\u4e60\u7684\u5165\u5ea6\u548c\u51fa\u5ea6\u5d4c\u5165\u3002\u7136\u540e\uff0c\u6ce8\u610f\u529b\u673a\u5236\u6709\u4e24\u4e2a\u504f\u7f6e\u9879\uff1a(1) \u8282\u70b9i\u548cj\u4e4b\u95f4\u6700\u77ed\u8def\u5f84\u7684\u8ddd\u79bb\uff1b(2) \u8fb9\u7279\u5f81\u7f16\u7801\uff0c\u5176\u53d6\u51b3\u4e8e\u53ef\u7528\u7684\u6700\u77ed\u8def\u5f84\u4e4b\u4e00\u3002<\/span><\/section>\n

      <\/section>\n<\/li>\n<\/ul>\n
      <\/figure>\n


      <\/p>\n

      \ud83c\udfc5 Graphormer \u5b8c\u6210\u4e86 2021 \u5e74 Graph ML \u7684\u5927\u6ee1\u8d2f\uff1a\u5728 OGB\u5927\u89c4\u6a21\u6311\u6218\u548c\u5f00\u653e\u50ac\u5316\u5242\u6311\u6218\u7684\u56fe\u56de\u5f52\u4efb\u52a1\u4e2d\u83b7\u5f97\u7b2c\u4e00\u540d\uff01<\/span><\/p>\n


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

      \ud83e\udd14 \u5f00\u653e\u6027\u95ee\u9898\uff1a\u53ef\u6269\u5c55\u6027\u548c\u8ba1\u7b97\u5f00\u9500\u3002SAN \u548c Graphormer \u5728\u5206\u5b50\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u5176\u4e2d\u56fe\u76f8\u5f53\u5c0f<\/span>\uff08\u5e73\u5747 50-100 \u4e2a\u8282\u70b9\uff09<\/span>\uff0c\u6211\u4eec\u53ef\u4ee5\u8d1f\u62c5\u5f97\u8d77\uff0c\u4f8b\u5982\uff0c\u8fd0\u884c O<\/span>(N\u00b3) <\/span>Floyd-Warshall \u6240\u6709\u5bf9\u6700\u77ed\u8def\u5f84\u3002\u6b64\u5916\uff0cGraph Transformers \u4ecd\u7136\u53d7\u5230 O<\/span>(N\u00b2) <\/span>\u6ce8\u610f\u529b\u673a\u5236\u7684\u9650\u5236\u3002\u5bf9\u4e8e\u5927\u56fe\u662f\u4e0d\u53ef\u63a5\u53d7\u7684\u3002\u60a8\u662f\u5426\u53ea\u662f\u60f3\u5230\u201cLinear transformers from NLP\u201d\uff1f\ud83d\ude09 \u662f\u7684\uff0c\u4ed6\u4eec\u53ef\u80fd\u4f1a\u4f38\u51fa\u63f4\u624b\uff0c\u4f46\u7531\u4e8e\u5176\u672a\u5b9e\u73b0\u6ce8\u610f\u529b\u77e9\u9635\uff0c\u56e0\u6b64\u9700\u8981\u627e\u5230\u4e00\u79cd\u806a\u660e\u7684\u65b9\u6cd5\u5c06\u8fb9\u7279\u5f81\u7f6e\u4e8e\u6b64\u7c7b\u6a21\u578b\u4e2d\u3002\u5f53\u7136\uff0c\u6211\u4eec\u4f1a\u5728 2022 \u5e74\u770b\u5230\u66f4\u591a\u5173\u4e8e\u8fd9\u65b9\u9762\u7684\u7814\u7a76\uff01<\/span><\/p>\n


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


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

      \n
      \n
      \n
      \n
      \n

      <\/section>\n

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

      2.<\/strong><\/span>Equivariant GNNs<\/span><\/strong><\/span><\/p>\n<\/section>\n

      \n

      <\/section>\n

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


      <\/p>\n

      <\/figure>\n


      <\/p>\n

      Geoffrey Hinton \u627f\u8ba4\u5176\u9177\u70ab\u7684\u7b49\u6548\u6027\u6709\u4f55\u72ec\u7279\u4e4b\u5904\uff1f<\/span><\/p>\n


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

      \u4e00\u822c\u6765\u8bf4\uff0cequivariance \u5b9a\u4e49\u5728\u67d0\u4e9b\u53d8\u6362\u7ec4\u4e0a\uff0c\u4f8b\u5982\uff0c3D \u65cb\u8f6c\u5f62\u6210 SO<\/span>(3)<\/span> \u7ec4\uff0c3 \u7ef4\u7279\u6b8a\u6b63\u4ea4\u7ec4\u3002equivariance \u6a21\u578b\u5728 2021 \u5e74\u6380\u8d77\u4e86 ML \u7684\u98ce\u66b4\ud83c\udf2a\uff0c\u5728 Graph ML \u4e2d\uff0c\u5b83\u5728\u8bb8\u591a\u5206\u5b50\u4efb\u52a1\u4e2d\u5c24\u5176\u5177\u6709\u7834\u574f\u6027\u3002\u5e94\u7528\u4e8e\u5206\u5b50\uff0cEquivariant GNN \u9700\u8981\u4e00\u4e2a\u989d\u5916\u7684\u8282\u70b9\u7279\u5f81\u8f93\u5165\u2014\u2014\u5373\u5206\u5b50\u7269\u7406\u5750\u6807\u7684\u4e00\u4e9b\u8868\u793a\uff0c\u8fd9\u4e9b\u8868\u793a\u5c06\u5728 n \u7ef4\u7a7a\u95f4\u4e2d\u65cb\u8f6c\/\u53cd\u5c04\/\u5e73\u79fb\u3002<\/span><\/p>\n


      <\/p>\n

      <\/figure>\n


      <\/p>\n

      \u4f7f\u7528\u7b49\u53d8\u6a21\u578b\uff0c\u5c3d\u7ba1\u53d8\u6362\u987a\u5e8f\u4e0d\u540c\uff0c\u6211\u4eec\u4ecd\u4f1a\u8fbe\u5230\u76f8\u540c\u7684\u6700\u7ec8\u72b6\u6001\u3002<\/span><\/p>\n


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

      Satorras\u3001 Hoogeboom\u548c Welling\u63d0\u51fa\u4e86EGNN\u3001<\/span>E(n) <\/span>\u7b49\u53d8 GNN\uff0c\u5176\u4e0e\u666e\u901a GNN \u7684\u91cd\u8981\u533a\u522b\u5728\u4e8e\u5411\u6d88\u606f\u4f20\u9012\u548c\u66f4\u65b0\u6b65\u9aa4\u6dfb\u52a0\u7269\u7406\u5750\u6807\u3002\u7b49\u5f0f 3 \ud83d\udc47\u5c06\u76f8\u5bf9\u5e73\u65b9\u8ddd\u79bb\u4e0e\u6d88\u606f m \u76f8\u52a0\uff0c\u7b49\u5f0f\u30024 \u66f4\u65b0\u4f4d\u7f6e\u7279\u5f81\u3002EGNN \u5728\u5efa\u6a21 n \u4f53\u7cfb\u7edf\u3001\u4f5c\u4e3a\u81ea\u52a8\u7f16\u7801\u5668\u548c\u91cf\u5b50\u5316\u5b66\u4efb\u52a1<\/span>\uff08QM9 \u6570\u636e\u96c6\uff09<\/span>\u65b9\u9762\u663e\u793a\u51fa\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u7ed3\u679c\u3002<\/span><\/p>\n


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

      <\/figure>\n

      <\/figure>\n
      \u4e0e\u666e\u901a GNN \u7684\u91cd\u8981\u533a\u522b\uff1a\u7b49\u5f0f 3 \u548c 4 \u5728\u6d88\u606f\u4f20\u9012\u548c\u66f4\u65b0\u6b65\u9aa4\u4e2d\u6dfb\u52a0\u4e86\u7269\u7406\u5750\u6807\u3002<\/span><\/section>\n

      <\/span><\/section>\n
      \u53e6\u4e00\u79cd\u9009\u62e9\u662f\u5728\u539f\u5b50\u4e4b\u95f4\u63ba\u5165\u89d2\u5ea6\u3002\u8fd9\u53ef\u80fd\u9700\u8981\u5c06\u8f93\u5165\u56fe\u8f6c\u6362\u4e3a\u7ebf\u6027\u56fe\uff0c\u5176\u4e2d\u6765\u81ea\u539f\u59cb\u56fe\u7684\u8fb9\u53d8\u6210\u7ebf\u6027\u56fe\u4e2d\u7684\u8282\u70b9\u3002\u8fd9\u6837\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5c06\u89d2\u5ea6\u4f5c\u4e3a\u65b0\u56fe\u4e2d\u7684\u8fb9\u7f18\u7279\u5f81\u3002<\/span><\/section>\n

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

      <\/section>\n


      <\/p>\n

      <\/span><\/h2>\n

      \n
      \n
      \n
      \n
      \n

      <\/section>\n

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

      3.<\/strong><\/span>Generative Models <\/span><\/strong><\/span><\/p>\n

      for Molecules<\/span><\/strong><\/span><\/p>\n<\/section>\n

      \n

      <\/section>\n

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


      <\/p>\n

      \u7531\u4e8e\u51e0\u4f55\u6df1\u5ea6\u5b66\u4e60\uff0c\u6574\u4e2a\u836f\u7269\u53d1\u73b0<\/span> (DD) <\/span>\u9886\u57df\u5728 2021 \u5e74\u5f97\u5230\u4e86\u663e\u8457\u7684\u63a8\u52a8\ud83d\ude80\u3002DD \u7684\u4f17\u591a\u5173\u952e\u6311\u6218\u4e4b\u4e00\u662f\u751f\u6210\u5177\u6709\u6240\u9700\u5c5e\u6027\u7684\u5206\u5b50<\/span>\uff08\u56fe\uff09<\/span>\u3002\u8fd9\u4e2a\u9886\u57df\u5f88\u5927\uff0c\u6240\u4ee5\u6211\u4eec\u53ea\u5f3a\u8c03\u6a21\u578b\u7684\u4e09\u4e2a\u5206\u652f\u3002<\/span><\/p>\n


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

        \n
      • \n

        \u89c4\u8303\u5316\u6d41\u7a0b\u3002Satorras\u3001Hoogeboom \u7b49\u4eba\u5e94\u7528\u4e0a\u8ff0\u7b49\u65b9\u5dee\u6846\u67b6\u6765\u521b\u5efaE(n) \u7b49\u53d8\u5f52\u4e00\u5316\u6d41\uff0c\u80fd\u591f\u751f\u6210\u5177\u6709\u4f4d\u7f6e\u548c\u7279\u5f81\u7684 3D \u5206\u5b50\ud83d\udcaa<\/span><\/p>\n<\/li>\n<\/ul>\n


        <\/p>\n

        <\/figure>\n


        <\/p>\n

          \n
        • \n

          \u6982\u7387\u6a21\u578b\u3002Shi\u3001Luo \u7b49\u4eba\u7814\u7a76\u4e86\u5728\u7ed9\u5b9a 2D \u56fe\u5f62\u7684\u60c5\u51b5\u4e0b\u751f\u6210 3D \u6784\u8c61\u5f02\u6784\u4f53<\/span>\uff08\u5373 3D \u7ed3\u6784\uff09<\/span>\u7684\u95ee\u9898\u3002\u6a21\u578bConfGF\u4f30\u8ba1\u539f\u5b50\u5750\u6807\u5bf9\u6570\u5bc6\u5ea6\u7684\u68af\u5ea6\u573a\u3002\u8fd9\u4e9b\u5b57\u6bb5\u662f\u65cb\u8f6c\u5e73\u79fb\u7b49\u53d8\u7684\uff0c\u4f5c\u8005\u60f3\u51fa\u4e86\u4e00\u79cd\u65b9\u6cd5\u5c06\u8fd9\u79cd\u7b49\u53d8\u5c5e\u6027\u5408\u5e76\u5230\u4f30\u8ba1\u5668\u4e2d\u3002Conformer \u91c7\u6837\u672c\u8eab\u662f\u901a\u8fc7\u9000\u706b\u52a8\u6001\u91c7\u6837\u5b8c\u6210\u7684\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n


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


          <\/p>\n

            \n
          • \n

            RL \u65b9\u6cd5\u3002\u4ee5\u4e00\u79cd\u975e\u5e38\u975e\u79d1\u5b66\u7684\u65b9\u5f0f\u63cf\u8ff0\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u901a\u8fc7\u9010\u6b65\u9644\u52a0\u201c\u6784\u5efa\u5757\u201d\u6765\u751f\u6210\u5206\u5b50\u3002\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u8fd9\u4e9b\u65b9\u6cd5\u5bf9\u6784\u5efa\u8fc7\u7a0b\u7684\u8c03\u8282\u65b9\u5f0f\u5bf9\u5176\u8fdb\u884c\u5e7f\u4e49\u5206\u7c7b\u3002\u4f8b\u5982\uff0cGao\u3001Mercado \u548c Coley\u4ee5\u53ef\u5408\u6210\u6027\u4e3a\u6784\u5efa\u8fc7\u7a0b\u7684\u6761\u4ef6\uff0c\u5373\u6211\u4eec\u662f\u5426\u53ef\u4ee5\u5728\u5b9e\u9a8c\u5ba4\u4e2d\u5b9e\u9645\u521b\u5efa \ud83e\uddea \u8fd9\u79cd\u5206\u5b50\u3002\u4e3a\u6b64\uff0c\u4ed6\u4eec\u9996\u5148\u5b66\u4e60\u5982\u4f55\u521b\u5efa\u6784\u5efa\u5757\u7684\u5408\u6210\u6811<\/span>\uff08\u4e00\u79cd\u6a21\u677f\uff09<\/span>\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n


            <\/p>\n

            <\/figure>\n

            <\/figure>\n
            \u7531 Yoshua Bengio \u9886\u5bfc\u7684 Mila \u548c\u65af\u5766\u798f\u7814\u7a76\u4eba\u5458\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u4e2a\u66f4\u901a\u7528\u7684\u6846\u67b6\uff0c\u4ed6\u4eec\u5f15\u5165\u4e86\u751f\u6210\u6d41\u7f51\u7edc<\/span>\uff08GFlowNets\uff09<\/span>\u3002\u5f88\u96be\u7528\u51e0\u53e5\u8bdd\u6765\u6982\u62ec\u2014\u2014\u9996\u5148\uff0c\u5f53\u6211\u4eec\u60f3\u8981\u5bf9\u4e0d\u540c\u7684\u5019\u9009\u8005\u8fdb\u884c\u91c7\u6837\u65f6\uff0cGFlowNets \u53ef\u4ee5\u7528\u4e8e\u4e3b\u52a8\u5b66\u4e60\u6848\u4f8b\uff0c\u5e76\u4e14\u91c7\u6837\u6982\u7387\u4e0e\u5956\u52b1\u51fd\u6570\u6210\u6b63\u6bd4\u3002\u6b64\u5916\uff0c\u4ed6\u4eec\u6700\u8fd1\u7684 NeurIPS’21 \u8bba\u6587\u5c55\u793a\u4e86 GFlowNets \u5e94\u7528\u4e8e\u5206\u5b50\u751f\u6210\u4efb\u52a1\u7684\u597d\u5904\u3002<\/span><\/section>\n

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

            <\/section>\n

            <\/span><\/h2>\n

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

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

            4.<\/strong>GNNs + Combinatorial <\/span><\/strong><\/span><\/p>\n

            Optimization & <\/span><\/strong><\/span>Algorithms<\/span><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

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


            <\/p>\n

            2021 \u5e74\u5bf9\u4e8e\u8fd9\u4e2a\u65b0\u5174\u7684\u5b50\u9886\u57df\u6765\u8bf4\u662f\u91cd\u8981\u7684\u4e00\u5e74\uff01Xu \u7b49\u4eba\u5728\u4ed6\u4eec\u6770\u51fa\u7684 ICLR’21 \u8bba\u6587\u4e2d\u7814\u7a76\u4e86\u795e\u7ecf\u7f51\u7edc\u7684\u5916\u63a8\uff0c\u5e76\u5f97\u51fa\u4e86\u51e0\u4e2a\u60ca\u4eba\u7684\u7ed3\u679c\u3002\u901a\u8fc7\u7b97\u6cd5\u5bf9\u9f50\u7684\u6982\u5ff5\uff0c\u4f5c\u8005\u8868\u660e GNN \u4e0e\u52a8\u6001\u89c4\u5212 <\/span>(DP)<\/span>\u5f88\u597d\u5730\u5bf9\u9f50<\/span>\uff08\u67e5\u770b\u63d2\u56fe \ud83d\udc47\uff09<\/span>\u3002\u4e8b\u5b9e\u4e0a\uff0c\u6bd4\u8f83\u7ecf\u5178 Bellman-Ford \u7b97\u6cd5\u7684\u8fed\u4ee3\u4ee5\u5bfb\u627e\u6700\u77ed\u8def\u5f84\u548c\u901a\u8fc7 GNN \u7684\u6d88\u606f\u7684\u805a\u5408\u7ec4\u5408\u6b65\u9aa4\u2014\u2014\u4f60\u4f1a\u53d1\u73b0\u5f88\u591a\u5171\u540c\u70b9\u3002\u6b64\u5916\uff0c\u4f5c\u8005\u8868\u660e\uff0c\u5728\u5efa\u6a21\u7279\u5b9a DP \u7b97\u6cd5\u65f6\uff0c\u4e3a GNN \u9009\u62e9\u5408\u9002\u7684\u805a\u5408\u51fd\u6570\u81f3\u5173\u91cd\u8981\uff0c\u4f8b\u5982\uff0c\u5bf9\u4e8e Bellman-Ford\uff0c\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u6700\u5c0f\u805a\u5408\u5668\u3002\u6211\u8fd8\u5efa\u8bae\u60a8\u67e5\u770b\u4e00\u4e2a\u975e\u5e38\u5177\u6709\u8bf4\u660e\u6027\u7684\u6f14\u8bb2 \u4f5c\u8005\uff1aStefanie Jegelka \u5728 2021 \u5e74\u6df1\u5ea6\u5b66\u4e60\u548c\u7ec4\u5408\u4f18\u5316\u7814\u8ba8\u4f1a\u4e0a\u89e3\u91ca\u4e86\u8fd9\u9879\u5de5\u4f5c\u7684\u4e3b\u8981\u6210\u679c\u3002<\/span><\/section>\n


            <\/p>\n

            <\/figure>\n


            <\/p>\n

            \u4e3a\u4e86\u66f4\u5168\u9762\u5730\u4ecb\u7ecd\u8be5\u9886\u57df\uff0c\u6211\u60f3\u91cd\u70b9\u4ecb\u7ecdCappart \u7b49\u4eba \u5728IJCAI’21 \u4e0a\u8fdb\u884c\u7684\u4e00\u9879\u5168\u9762\u8c03\u67e5\uff0c\u8be5\u8c03\u67e5\u6db5\u76d6\u4e86\u7ec4\u5408\u4f18\u5316\u4e2d\u7684 GNN\u3002\u8fd9\u7bc7\u6587\u7ae0\u9996\u6b21\u51fa\u73b0\u4e86\u795e\u7ecf\u7b97\u6cd5\u63a8\u7406\u84dd\u56fe\u3002<\/span><\/p>\n


            <\/p>\n

            <\/p>\n


            <\/p>\n

            \ud83d\udcd8 \u84dd\u56fe\u89e3\u91ca\u4e86\u795e\u7ecf\u7f51\u7edc\u5982\u4f55\u5728\u5d4c\u5165\u7a7a\u95f4\u4e2d\u6a21\u4eff\u548c\u6388\u6743\u901a\u5e38\u79bb\u6563\u7b97\u6cd5\u7684\u6267\u884c\u8fc7\u7a0b\u3002\u5728\u7f16\u7801-\u5904\u7406-\u89e3\u7801\u65b9\u5f0f\u4e2d\uff0c\u62bd\u8c61\u8f93\u5165<\/span>\uff08\u4ece\u81ea\u7136\u8f93\u5165\u83b7\u5f97\uff09<\/span>\u7531\u795e\u7ecf\u7f51\u7edc<\/span>\uff08\u5904\u7406\u5668\uff09<\/span>\u5904\u7406\uff0c\u5176\u8f93\u51fa\u88ab\u89e3\u7801\u4e3a\u62bd\u8c61\u8f93\u51fa\uff0c\u7136\u540e\u53ef\u4ee5\u6620\u5c04\u5230\u66f4\u81ea\u7136\u7684\u4efb\u52a1\u7279\u5b9a\u8f93\u51fa\u3002\u4f8b\u5982\uff0c\u5982\u679c\u62bd\u8c61\u8f93\u5165\u548c\u8f93\u51fa\u53ef\u4ee5\u8868\u793a\u4e3a\u56fe\u5f62\uff0c\u90a3\u4e48 GNN \u53ef\u4ee5\u662f\u5904\u7406\u5668\u7f51\u7edc\u3002\u7684\u79bb\u6563\u7b97\u6cd5\u7684\u4e00\u4e2a\u5e38\u89c1\u7684\u9884\u5904\u7406\u6b65\u9aa4\u662f\u58c1\u7403\u65e0\u8bba\u6211\u4eec\u4e86\u89e3\u8fd9\u4e2a\u95ee\u9898\uff0c\u5230\u50cf\u201c\u7684\u6807\u91cf\u8ddd\u79bb\u201d\u6216\u201c\u8fb9\u7f18\u5bb9\u91cf’ \u5e76\u5728\u8fd9\u4e9b\u6807\u91cf\u4e0a\u8fd0\u884c\u7b97\u6cd5\u3002\u76f8\u53cd\uff0c\u5411\u91cf\u8868\u793a\u548c\u795e\u7ecf\u6267\u884c\u53ef\u4ee5\u8f7b\u677e\u542f\u7528\u9ad8\u7ef4\u8f93\u5165\u800c\u4e0d\u662f\u7b80\u5355\u7684\u6807\u91cf\uff0c\u5e76\u9644\u52a0\u53cd\u5411\u4f20\u64ad\u4ee5\u4f18\u5316\u5904\u7406\u5668\u3002<\/span><\/p>\n


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


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

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

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

            5.<\/strong><\/span>Subgraph GNNs:<\/span><\/strong><\/span>Beyond 1-WL<\/span><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

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


            <\/p>\n

            Subgraph GNNs: Beyond 1-WL<\/span><\/h2>\n

            \ud83e\udd1d \u5982\u679c\u8bf4 2020 \u5e74\u662f\u7b2c\u4e00\u6b21\u5c1d\u8bd5\u79bb\u5f00 GNN \u8868\u73b0\u529b\u7684 1-WL-landia \u7684\u4e00\u5e74\uff0c\u90a3\u4e48\u6211\u4eec\u53ef\u4ee5\u786e\u8ba4 2021 \u5e74\u662f\u4e0e\u8d85\u8d8a 1WL-landia \u7684\u5916\u661f\u4eba\u5efa\u7acb\u7a33\u56fa\u8054\u7cfb\u7684\u4e00\u5e74\ud83d\udc7e\u3002<\/span><\/p>\n


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

            \u5982\u679c\u60a8\u60f3\u771f\u6b63\u6df1\u5165\u7814\u7a76 WL \u7684\u6700\u65b0\u7814\u7a76\uff0c\u8bf7\u67e5\u770b\u7531 Christopher Morris\u3001Yaron Lipman\u3001Haggai Maron \u548c\u5408\u8457\u8005\u64b0\u5199\u7684\u6700\u65b0\u8c03\u67e5Weisfeiler \u548c Leman go Machine Learning: The Story so far\u3002<\/span><\/p>\n


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


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

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

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

            6.<\/strong><\/span>Scalability and Deep GNNs: <\/span><\/strong><\/span><\/p>\n

            100 Layers <\/span><\/strong><\/span>and More<\/span><\/strong><\/p>\n<\/section>\n

            \n

            <\/section>\n

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


            <\/p>\n

            Scalability and Deep GNNs: 100 Layers and More<\/span><\/h2>\n


            <\/p>\n

            \u5982\u679c\u60a8\u5728\u4f7f\u7528 2-4 \u5c42 GNN \u65f6\u5ac9\u5992\u6df1\u5ea6 ResNets \u6216 100 \u5c42\u4ee5\u4e0a\u7684\u5de8\u5927 Transformer\uff0c\u90a3\u4e48\u662f\u65f6\u5019\u6b22\u547c\u4e86\ud83e\udd29\uff012021 \u5e74\u4e3a\u6211\u4eec\u5e26\u6765\u4e86 2 \u7bc7\u968f\u610f\u8bad\u7ec3100-1000 \u5c42GNN \u7684\u8bba\u6587\uff0c\u4ee5\u53ca\u4e00\u7bc7\u5173\u4e8e\u51e0\u4e4e\u6052\u5b9a\u5927\u5c0f\u7684\u90bb\u57df\u91c7\u6837\u7684\u5de5\u4f5c\u3002Li \u7b49\u4eba\u63d0\u51fa\u4e86\u4e24\u79cd\u673a\u5236\uff0c\u5728\u8bad\u7ec3\u6781\u6df1\u7684\u8d85\u53c2\u6570\u5316\u7f51\u7edc\u65f6\uff0c\u5c06 GPU \u5185\u5b58\u6d88\u8017\u4eceL \u5c42\u7684O<\/span>(L)<\/span>\u663e\u7740\u964d\u4f4e\u5230O<\/span>(1)<\/span>\u3002\u4f5c\u8005\u5c55\u793a\u4e86\u5982\u4f55\u4f7f\u7528 1\ufe0f\u20e3 \u53ef\u9006\u5c42\u5728 CV \u6216\u9ad8\u6548\u7684 Transformer \u67b6\u6784<\/span>\uff08\u5982Reformer \uff09<\/span>\u4e2d\u4f7f\u7528\u591a\u5e74\uff1b2\ufe0f\u20e3 \u5728\u5c42\u4e4b\u95f4\u5171\u4eab\u6743\u91cd<\/span>\uff08\u6743\u91cd\u7ed1\u5b9a\uff09<\/span>\u3002\u7136\u540e\u53ef\u4ee5\u8bad\u7ec3\u591a\u8fbe 1000 \u5c42\u7684 GNN \ud83d\udc40\u3002\u4e0b\u9762\u7684\u56fe\u8868\u5c55\u793a\u4e86\u5bf9 GPU \u8981\u6c42\u9002\u4e2d\u7684\u5c42\u6570\u7684\u4e0d\u65ad\u6269\u5c55\u3002<\/span><\/p>\n


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

            <\/figure>\n


            <\/p>\n

            Godwin \u7b49\u4eba\u4ecb\u7ecd\u4e86\u4e00\u79cd\u5229\u7528\u5faa\u73af\u5b66\u4e60\u6df1\u5ea6 GNN \u7684\u65b9\u6cd5\u2014\u2014\u6d88\u606f\u4f20\u9012\u6b65\u9aa4\u7ec4\u7ec7\u5728\u5757\u4e2d\uff0c\u6bcf\u4e2a\u5757\u53ef\u4ee5\u6709 M \u4e2a\u6d88\u606f\u4f20\u9012\u5c42\u3002\u7136\u540e\uff0c\u5faa\u73af\u5e94\u7528 N \u4e2a\u5757\uff0c\u8fd9\u610f\u5473\u7740\u5757\u5171\u4eab\u6743\u91cd\u3002\u5982\u679c\u60a8\u6709 10 \u4e2a\u6d88\u606f\u4f20\u9012\u5c42\u548c 10 \u4e2a\u5757\uff0c\u60a8\u5c06\u5f97\u5230\u4e00\u4e2a 100 \u5c42\u7684 GNN\u3002\u4e00\u4e2a\u91cd\u8981\u7684\u7ec4\u6210\u90e8\u5206\u662f\u566a\u58f0\u8282\u70b9\u6b63\u5219\u5316\u6280\u672f\uff0c\u5b83\u6270\u52a8\u8282\u70b9\u548c\u8fb9\u7f18\u7279\u5f81\u5e76\u8ba1\u7b97\u989d\u5916\u7684\u53bb\u566a\u635f\u5931\u3002\u8be5\u67b6\u6784\u9002\u5408\u66f4\u597d\u7684\u5206\u5b50\u4efb\u52a1\uff0c\u5e76\u5728 QM9 \u548c OpenCatalyst20 \u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\u3002<\/span><\/p>\n


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

            <\/figure>\n


            <\/p>\n

            \u6700\u540e\uff0c\u5982\u679c\u6211\u4eec\u60f3\u5c06\u4efb\u610f GNN \u7f29\u653e\u5230\u975e\u5e38\u5927\u7684\u56fe\uff0c\u9664\u4e86\u91c7\u6837\u5b50\u56fe\u4e4b\u5916\uff0c\u6211\u4eec\u522b\u65e0\u9009\u62e9\u3002\u901a\u5e38\uff0c\u5bf9 k \u8df3\u5b50\u56fe\u8fdb\u884c\u91c7\u6837\u4f1a\u5bfc\u81f4\u6307\u6570\u7ea7\u5185\u5b58\u6210\u672c\u548c\u8ba1\u7b97\u56fe\u5927\u5c0f\u3002<\/span><\/p>\n


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

            PyG \u7684\u4f5c\u8005Matthias Fey \u7b49\u4eba\u521b\u5efa\u4e86GNNAutoScale\uff0c\u8fd9\u662f\u4e00\u4e2a\u5728\u6052\u5b9a\u65f6\u95f4\u5185\u5229\u7528\u5386\u53f2\u5d4c\u5165<\/span>\uff08\u6765\u81ea\u5148\u524d\u6d88\u606f\u4f20\u9012\u6b65\u9aa4\u7684\u7f13\u5b58\u7684\u5947\u7279\u7248\u672c\uff09<\/span>\u548c\u56fe\u805a\u7c7b<\/span>\uff08\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u662f\u4f17\u6240\u5468\u77e5\u7684 METIS \u7b97\u6cd5\uff09<\/span>\u6269\u5c55 GNN \u7684\u6846\u67b6. \u5728\u9884\u5904\u7406\u671f\u95f4\uff0c\u6211\u4eec\u5c06\u56fe\u5212\u5206\u4e3a B \u4e2a\u96c6\u7fa4<\/span>\uff08\u5c0f\u6279\u91cf\uff09<\/span>\uff0c\u4ee5\u4fbf\u6700\u5c0f\u5316\u96c6\u7fa4\u4e4b\u95f4\u7684\u8fde\u901a\u6027\u3002\u7136\u540e\u6211\u4eec\u53ef\u4ee5\u5728\u8fd9\u4e9b\u96c6\u7fa4\u4e0a\u8fd0\u884c\u6d88\u606f\u4f20\u9012\u4ee5\u8ddf\u8e2a\u7f13\u5b58\u4e2d\u66f4\u65b0\u7684\u8282\u70b9\u529f\u80fd\u3002\u5728\u5b9e\u9a8c\u4e0a\uff0c\u6df1\u5ea6\u7f51\u7edc\u7684 GNNAutoScale<\/span>\uff08\u6700\u591a 64 \u5c42\uff09<\/span>\u7684\u6027\u80fd\u4e0e\u5168\u6279\u6b21\u8bbe\u7f6e\u4e00\u6837\u597d\uff0c\u4f46\u663e\u7740\u964d\u4f4e\u4e86\u5185\u5b58\u8981\u6c42<\/span>\uff08\u5c0f\u4e86\u7ea6 50 \u500d\uff09<\/span>\u2014\u2014\u56e0\u6b64\u60a8\u53ef\u4ee5\u5728\u5546\u54c1\u7ea7 GPU \u4e0a\u5b89\u88c5\u6df1\u5ea6 GNN \u548c\u5927\u56fe \ud83d\udcaa<\/span><\/p>\n


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

            <\/figure>\n


            <\/p>\n


            <\/p>\n

            \n
            \n
            \n
            \n
            \n

            <\/section>\n

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

            7.<\/strong><\/span>Knowledge Graph<\/span><\/strong><\/span><\/p>\n<\/section>\n

            \n

            <\/section>\n

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


            <\/p>\n

            KGs \u4e0a\u7684\u8868\u5f81\u5b66\u4e60\u7ec8\u4e8e\u7a81\u7834\u4e86\u8f6c\u5bfc\u7684\u5929\u82b1\u677f\u3002\u5728 2021 \u5e74\u4e4b\u524d\uff0c\u6a21\u578b\u660e\u786e\u5206\u4e3a\u8f6c\u5bfc\u548c\u5f52\u7eb3\uff0c\u5177\u6709\u4e0d\u540c\u7684\u5f52\u7eb3\u504f\u5dee\u3001\u67b6\u6784\u548c\u8bad\u7ec3\u673a\u5236\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u8f6c\u5bfc\u6a21\u578b\u6ca1\u6709\u673a\u4f1a\u9002\u5e94\u770b\u4e0d\u89c1\u7684\u5b9e\u4f53\uff0c\u800c\u5f52\u7eb3\u6a21\u578b\u5728\u4e2d\u5927\u578b\u56fe\u4e0a\u8bad\u7ec3\u6210\u672c\u592a\u9ad8\u30022021 \u5e74\u5e26\u6765\u4e86\u4e24\u79cd\u67b6\u6784<\/span><\/p>\n


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

              \n
            • \n

              \u5728\u8f6c\u5bfc\u548c\u5f52\u7eb3\u73af\u5883\u4e2d\u5de5\u4f5c\uff0c<\/span><\/p>\n<\/li>\n

            • \n

              \u4e0d\u9700\u8981\u8282\u70b9\u7279\u5f81\uff0c<\/span><\/p>\n<\/li>\n

            • \n

              \u53ef\u4ee5\u5728\u5f52\u7eb3\u6a21\u5f0f\u4e2d\u4ee5\u4e0e\u8f6c\u5bfc\u6a21\u5f0f\u76f8\u540c\u7684\u65b9\u5f0f\u8fdb\u884c\u8bad\u7ec3\uff0c<\/span><\/p>\n<\/li>\n

            • \n

              \u53ef\u6269\u5c55\u5230\u73b0\u5b9e\u4e16\u754c\u7684 KG \u5927\u5c0f\u3002\u7b2c\u4e00\u4e2a\u662fNeural Bellman-Ford nets\u3002\u4f5c\u8005\u627e\u5230\u4e86\u4e00\u79cd\u975e\u5e38\u4f18\u96c5\u7684\u65b9\u6cd5\u5c06\u7ecf\u5178\u7684 Bellman-Ford \u63a8\u5e7f\u5230\u66f4\u9ad8\u7ea7\u522b\u7684\u6846\u67b6\uff0c\u5e76\u5c55\u793a\u4e86\u6211\u4eec\u5982\u4f55\u901a\u8fc7\u4f7f\u7528\u7279\u5b9a\u8fd0\u7b97\u7b26\u5b9e\u4f8b\u5316\u6846\u67b6\u6765\u83b7\u5f97\u5176\u4ed6\u77e5\u540d\u65b9\u6cd5\uff0c\u4f8b\u5982Katz index\u3001PPR\u6216Widest Path\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u4ed6\u4eec\u8868\u660e\u5e7f\u4e49 Bellman-Ford \u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u5173\u7cfb GNN \u67b6\u6784<\/span>\uff08\u7b97\u6cd5\u5bf9\u9f50\u7684\u53e6\u4e00\u4e2a\u786e\u8ba4GNN \u548c\u52a8\u6001\u89c4\u5212\u4e4b\u95f4\uff09<\/span>\u3002NBFNet \u4e0d\u5b66\u4e60\u5b9e\u4f53\u5d4c\u5165<\/span>\uff08\u4ec5\u5173\u7cfb\u548c GNN \u6743\u91cd\uff09<\/span>\uff0c\u8fd9\u4f7f\u6a21\u578b\u901a\u8fc7\u8bbe\u8ba1\u548c\u6cdb\u5316\u5230\u770b\u4e0d\u89c1\u7684\u56fe\u5177\u6709\u5f52\u7eb3\u6027\u3002\u8be5\u6a21\u578b\u5728\u5173\u7cfb\u56fe\u548c\u975e\u5173\u7cfb\u56fe\u4e0a\u7684\u94fe\u63a5\u9884\u6d4b\u4efb\u52a1\u4e0a\u90fd\u8868\u73b0\u51fa\u8272\u3002\u5e94\u7528\u4e8e KG\uff0cNBFNet \u4ece 2019 \u5e74\u5f00\u59cb\u4e3a FB15k-237 \u548c WN18RR \u5e26\u6765\u6700\u5927\u7684\u6027\u80fd\u63d0\u5347\uff0c\u540c\u65f6\u53c2\u6570\u51cf\u5c11\u4e86 100 \u500d\ud83d\udcaa\u3002<\/span><\/p>\n<\/li>\n<\/ul>\n


              <\/p>\n

              <\/figure>\n


              <\/p>\n

              \ud83c\udf70 Galkin \u7b49\u4eba\u7684\u53e6\u4e00\u79cd\u65b9\u6cd5<\/span>\uff08\u514d\u8d23\u58f0\u660e\uff1a\u672c\u6587\u7684\u4f5c\u8005\u4e4b\u4e00\u662f\u8bba\u6587\u7684\u4f5c\u8005\uff09<\/span>\u7684\u7075\u611f\u6765\u81ea NLP \u4e2d\u7684\u6807\u8bb0\u5316\u7b97\u6cd5\uff0c\u8be5\u7b97\u6cd5\u5177\u6709\u56fa\u5b9a\u7684\u6807\u8bb0\u8bcd\u6c47\u8868\uff0c\u80fd\u591f\u6807\u8bb0\u4efb\u4f55\u5355\u8bcd\uff0c\u751a\u81f3\u662f\u90a3\u4e9b\u5728\u8bad\u7ec3\u671f\u95f4\u770b\u4e0d\u89c1\u7684\u5355\u8bcd\u65f6\u95f4\u3002\u5e94\u7528\u4e8eKG\uff0cNodePiece\u5c06\u6bcf\u4e2a\u8282\u70b9\u8868\u793a\u4e3a\u4e00\u7ec4top-k\u6700\u8fd1\u7684\u951a\u8282\u70b9<\/span>\uff08\u53ef\u9009\u5730\u5728\u9884\u5904\u7406\u6b65\u9aa4\u4e2d\u91c7\u6837\uff09<\/span>\u548cm\u8282\u70b9\u5468\u56f4\u7684\u552f\u4e00\u5173\u7cfb\u7c7b\u578b\u3002\u951a\u70b9\u548c\u5173\u7cfb\u7c7b\u578b\u88ab\u7f16\u7801\u4e3a\u53ef\u7528\u4e8e\u4efb\u4f55\u4e0b\u6e38\u4efb\u52a1<\/span>\uff08\u5206\u7c7b\u3001\u94fe\u63a5\u9884\u6d4b\u3001\u5173\u7cfb\u9884\u6d4b\uff0c\u4ec5\u4e3e\u51e0\u4f8b\uff09<\/span>\u548c\u4efb\u4f55\u5f52\u7eb3\/\u8f6c\u5bfc\u8bbe\u7f6e\u7684\u8282\u70b9\u8868\u793a\u3002NodePiece \u7279\u5f81\u53ef\u4ee5\u76f4\u63a5\u88ab RotatE \u7b49\u975e\u53c2\u6570\u89e3\u7801\u5668\u4f7f\u7528\uff0c\u4e5f\u53ef\u4ee5\u53d1\u9001\u5230 GNN \u8fdb\u884c\u6d88\u606f\u4f20\u9012\u3002\u8be5\u6a21\u578b\u5728\u5f52\u7eb3\u94fe\u63a5\u9884\u6d4b\u6570\u636e\u96c6\u4e0a\u4e0e NBFNet \u76f8\u6bd4\u5177\u6709\u7ade\u4e89\u529b\uff0c\u5e76\u5728\u5927\u56fe\u4e0a\u8868\u73b0\u51fa\u9ad8\u53c2\u6570\u6548\u7387\u2014\u2014OGB WikiKG 2\u4e0a\u7684 NodePiece \u6a21\u578b\u9700\u8981\u7684\u53c2\u6570\u6bd4\u6d45\u5c42\u8f6c\u5bfc\u6a21\u578b\u5c11\u7ea6 100 \u500d\u3002<\/span><\/p>\n


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

              <\/figure>\n


              <\/h2>\n


              <\/p>\n

              \n
              \n
              \n
              \n
              \n

              <\/section>\n

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

              8.<\/strong><\/span>Generally Cool Research <\/span><\/strong><\/span><\/p>\n

              with GNNs<\/span><\/strong><\/span><\/p>\n<\/section>\n

              \n

              <\/section>\n

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


              <\/p>\n

              \u672c\u8282\u63d0\u5230\u4e86\u51e0\u4e2a\u7279\u522b\u9177\u7684\u4f5c\u54c1\uff0c\u5b83\u4eec\u4f7f\u7528\u4e86 GNN\uff0c\u4f46\u4e0d\u5c5e\u4e8e\u67d0\u4e2a\u7279\u5b9a\u7c7b\u522b\u3002Huang\u3001He \u7b49\u4eba\u5728 ICLR’21 \u4e0a\u5c55\u793a\u4e86Correct & Smooth \u2014 \u4e00\u4e2a\u901a\u8fc7\u6807\u7b7e\u4f20\u64ad\u6539\u8fdb\u6a21\u578b\u9884\u6d4b\u7684\u7b80\u5355\u7a0b\u5e8f\u3002\u4ec5\u4e0e MLP \u642d\u914d\u4f7f\u7528\uff0c\u8be5\u65b9\u6cd5\u5728\u4e0d\u4f7f\u7528\u4efb\u4f55 GNN \u4e14\u53c2\u6570\u5c11\u5f97\u591a\u7684\u60c5\u51b5\u4e0b\u4ee5\u6700\u9ad8\u5206\u51b2\u51fb OGB \u6392\u884c\u699c\uff01\u4eca\u5929\uff0c\u51e0\u4e4e\u6240\u6709OGB \u8282\u70b9\u5206\u7c7b\u8f68\u9053\u4e2d\u7684\u9876\u7ea7\u6a21\u578b\u90fd\u4f7f\u7528 Correct & Smooth \u6765\u6324\u538b\u66f4\u591a\u7684\u70b9\u3002<\/span><\/p>\n


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

              <\/figure>\n


              <\/p>\n

              \ud83e\ude84 Knyazev \u7b49\u4eba\u5728 11 \u6708\u4ee5\u4e00\u6b21\u524d\u5411\u4f20\u9012\u4e2d\u9884\u6d4b\u5404\u79cd\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\u53c2\u6570\u7684\u5de5\u4f5c\u9707\u60ca\u4e86 ML \u793e\u533a\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u4e0d\u662f\u968f\u673a\u521d\u59cb\u5316\u6a21\u578b\uff0c\u800c\u662f\u53ef\u4ee5\u7acb\u5373\u9884\u6d4b\u51fa\u597d\u7684\u53c2\u6570\uff0c\u8fd9\u6837\u7684\u6a21\u578b\u5df2\u7ecf\u6bd4\u968f\u673a\u6a21\u578b\u597d\u5f97\u591a\u4e86\ud83d\udc40\u3002\u5f53\u7136\uff0c\u5982\u679c\u4f60\u7528n \u4e2aSGD \u6b65\u9aa4\u4f18\u5316\u4e00\u4e2a\u968f\u673a\u521d\u59cb\u5316\u7684\u7f51\u7edc\uff0c\u4f60\u4f1a\u5f97\u5230\u66f4\u9ad8\u7684\u6570\u5b57\uff0c\u4f46\u672c\u6587\u7684\u4e00\u4e2a\u4e3b\u8981\u8d21\u732e\u662f\uff0c\u901a\u5e38\u53ef\u4ee5\u5728\u4e0d \u8bad\u7ec3\u8fd9\u4e2a\u7279\u5b9a\u67b6\u6784\u7684\u60c5\u51b5\u4e0b\u627e\u5230\u5408\u9002\u7684\u53c2\u6570\u3002\u53c2\u6570\u9884\u6d4b\u5b9e\u9645\u4e0a\u662f\u4e00\u4e2a\u56fe\u5b66\u4e60\u4efb\u52a1\u2014\u2014\u4efb\u4f55\u795e\u7ecf\u7f51\u7edc\u67b6\u6784<\/span>\uff08ResNet\u3001ViT\u3001Transformers\uff0c\u4f60\u8bf4\u7684\uff09<\/span>\u90fd\u53ef\u4ee5\u8868\u793a\u4e3a\u4e00\u4e2a\u8ba1\u7b97\u56fe\uff0c\u5176\u4e2d\u8282\u70b9\u662f\u5177\u6709\u53ef\u5b66\u4e60\u53c2\u6570\u7684\u6a21\u5757\uff0c\u8282\u70b9\u7279\u5f81\u662f\u90a3\u4e9b\u53c2\u6570\uff0c\u6211\u4eec\u6709\u4e00\u5806\u8282\u70b9\u7c7b\u578b<\/span>\uff08\u6bd4\u5982\uff0c\u7ebf\u6027\u5c42\u3001\u5377\u79ef\u5c42\u3001\u6279\u8303\u6570\uff0c\u4f5c\u8005\u4f7f\u7528\u4e86\u5927\u7ea6 15 \u79cd\u8282\u70b9\u7c7b\u578b\uff09<\/span>\u3002\u53c2\u6570\u9884\u6d4b\u5219\u662f\u4e00\u4e2a\u8282\u70b9\u56de\u5f52\u4efb\u52a1\u3002\u8ba1\u7b97\u56fe\u4f7f\u7528 GatedGNN \u8fdb\u884c\u7f16\u7801\uff0c\u5e76\u5c06\u5176\u65b0\u8868\u793a\u53d1\u9001\u5230\u89e3\u7801\u5668\u6a21\u5757\u3002\u4e3a\u4e86\u8bad\u7ec3\uff0c\u4f5c\u8005\u6536\u96c6\u4e86\u4e00\u4e2a\u5305\u542b 1M \u4e2a\u67b6\u6784<\/span>\uff08\u56fe\uff09<\/span>\u7684\u65b0\u6570\u636e\u96c6\u3002\u8be5\u65b9\u6cd5\u9002\u7528\u4e8e\u4efb\u4f55\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff0c\u751a\u81f3\u9002\u7528\u4e8e\u5176\u4ed6 GNN\uff01<\/span><\/p>\n


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

              \ud83d\uddfa DeepMind \u548c\u8c37\u6b4c\u901a\u8fc7\u5c06\u9053\u8def\u7f51\u7edc\u5efa\u6a21\u4e3a\u8d85\u5206\u6bb5\u56fe\u5e76\u5728\u5176\u4e0a\u5e94\u7528 GNN\uff0c\u6781\u5927\u5730\u63d0\u9ad8\u4e86\u8c37\u6b4c\u5730\u56fe\u4e2d ETA \u7684\u8d28\u91cf\u3002\u5728Pinion \u7b49\u4eba\u7684\u8bba\u6587\u4e2d\uff0c\u8be5\u4efb\u52a1\u88ab\u5b9a\u4e49\u4e3a\u8282\u70b9\u7ea7\u548c\u56fe\u7ea7\u56de\u5f52\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u4f5c\u8005\u8fd8\u63cf\u8ff0\u4e86\u8bb8\u591a\u9700\u8981\u89e3\u51b3\u7684\u5de5\u7a0b\u6311\u6218\uff0c\u4ee5\u4fbf\u5728\u8c37\u6b4c\u5730\u56fe\u89c4\u6a21\u4e0a\u90e8\u7f72\u7cfb\u7edf\u3002\u5e94\u7528 GNN \u89e3\u51b3\u6570\u767e\u4e07\u7528\u6237\u9762\u4e34\u7684\u5b9e\u9645\u95ee\u9898\u7684\u5b8c\u7f8e\u793a\u4f8b\uff01<\/span><\/p>\n


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

              <\/figure>\n


              <\/p>\n

              Transgene \u548c NEC\u6700\u8fd1\u5ba3\u5e03\u4e86GNN \u5728\u764c\u75c7\u7814\u7a76\u4e2d\u7684\u53e6\u4e00\u4e2a\u6f5c\u5728\u5f71\u54cd\u5e94\u7528\u3002\u6839\u636eNEC \u9996\u5e2d\u7814\u7a76\u5458Mathias Niepert \u7684\u8bf4\u6cd5\uff0cGNN \u88ab\u7528\u4e8e\u901a\u8fc7\u5d4c\u5165\u4f20\u64ad\u6765\u4f30\u7b97\u4e22\u5931\u7684\u60a3\u8005\u6570\u636e<\/span>\uff08\u67e5\u770b\u6700\u8fd1\u5173\u4e8e\u6b64\u65b9\u6cd5\u7684Twitter \u7ebf\u7a0b\uff09<\/span>\u4ee5\u53ca\u5bf9\u5019\u9009\u80bd\u8fdb\u884c\u8bc4\u5206\u4ee5\u5f15\u53d1\u514d\u75ab\u53cd\u5e94\u3002\u6700\u540e\uff0cDeepMind \u7684Davies \u7b49\u4eba\u6700\u8fd1\u4f7f\u7528 GNN \u6765\u5e2e\u52a9\u5236\u5b9a\u4e0e\u6838\u5fc3\u6570\u5b66\u95ee\u9898\u6709\u5173\u7684\u731c\u60f3<\/span>\uff08\u5173\u4e8e \ud83e\udea2 \u7ed3 \ud83e\udea2 \u5728\u4ed6\u4eec\u7684\u5de5\u4f5c\u4e2d\uff09<\/span>\uff0c\u5b9e\u9645\u4e0a\u786e\u5b9e\u627e\u5230\u5e76\u8bc1\u660e\u4e86\u4e00\u4e2a\u65b0\u5b9a\u7406\uff01\u4f60\u770b\uff0cGNN \u4e5f\u53ef\u4ee5\u5904\u7406\u975e\u5e38\u62bd\u8c61\u7684\u4e8b\u60c5 \ud83d\udc4f<\/span><\/p>\n


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


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

              \n
              \n
              \n
              \n
              \n

              <\/section>\n

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

              9.<\/strong><\/span>New Datasets, Challenges,<\/span><\/strong><\/span><\/p>\n

              and Tasks<\/span><\/strong><\/span><\/p>\n<\/section>\n

              \n

              <\/section>\n

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


              <\/p>\n

              \u5982\u679c\u60a8\u538c\u5026\u4e86 Cora\u3001Citeseer \u548c Pubmed \u2014 \u6211\u4eec\u4f1a\u611f\u89c9\u5230\u60a8\u30022021 \u5e74\u5e26\u6765\u4e86\u5927\u91cf\u4e0d\u540c\u5927\u5c0f\u548c\u7279\u5f81\u7684\u65b0\u6570\u636e\u96c6\u3002<\/span><\/p>\n


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