{"id":37318,"date":"2022-09-15T22:30:40","date_gmt":"2022-09-15T14:30:40","guid":{"rendered":"https:\/\/swarma.org\/?p=37318"},"modified":"2022-09-15T22:30:40","modified_gmt":"2022-09-15T14:30:40","slug":"%e9%9d%a2%e5%90%91%e5%9b%a0%e6%9e%9c%e8%a7%84%e5%be%8b%e7%9a%84%e8%a1%a8%e7%a4%ba%e5%ad%a6%e4%b9%a0%e6%96%b0%e6%96%b9%e6%b3%95-%e5%9b%a0%e6%9e%9c%e8%a1%a8%e5%be%81%e5%ad%a6%e4%b9%a0","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=37318","title":{"rendered":"\u9762\u5411\u56e0\u679c\u89c4\u5f8b\u7684\u8868\u793a\u5b66\u4e60\u65b0\u65b9\u6cd5\u2014\u2014\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u6700\u65b0\u653b\u7565"},"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

\u6211\u4eec\u7684\u751f\u6d3b\u4e2d\u65e0\u65f6\u65e0\u523b\u4e0d\u5728\u5bf9\u63a5\u53d7\u5230\u7684\u4fe1\u606f\u8fdb\u884c\u601d\u8003\u548c\u903b\u8f91\u63a8\u7406\uff0c\u4ece\u800c\u83b7\u5f97\u53ef\u88ab\u590d\u73b0\u7684\u6a21\u5757\u5316\u77e5\u8bc6\u3002\u8fd1\u5e74\u6765\u6df1\u5ea6\u5b66\u4e60AI\u5728\u9884\u6d4b\u8bc6\u522b\u65b9\u9762\u53d6\u5f97\u5de8\u5927\u8fdb\u5c55\u3002\u8981\u5982\u4f55\u8fdb\u4e00\u6b65\uff0c\u4f7fAI\u5177\u6709\u4e0e\u4eba\u7c7b\u76f8\u4f3c\u7684\u63a8\u7406\u80fd\u529b\u5462\uff1f\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u6216\u8bb8\u63d0\u4f9b\u4e86\u4e00\u6761\u8def\u5f84\u3002\u8fd9\u7bc7\u6587\u7ae0\u6982\u62ec\u4e86 Carnegie Mellon University \u526f\u6559\u6388\u5f20\u5764\u8001\u5e08\u7814\u7a76\u7ec4\u5728\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u65b9\u9762\u7684\u6700\u65b0\u5de5\u4f5c\uff08\u5f20\u5764\u76ee\u524d\u5b66\u672f\u4f11\u5047\uff0c\u5728MBZUAI\u5de5\u4f5c\uff09\u3002\u5f20\u5764\u56e2\u961f\u81f4\u529b\u4e8e\u6253\u901a\u56e0\u679c\u5173\u7cfb\u548c\u4eba\u5de5\u667a\u80fd\u7684\u5173\u7cfb\uff1a\u4e00\u65b9\u9762\uff0c\u4ed6\u4eec\u5f00\u53d1\u673a\u5668\u5b66\u4e60\u65b9\u6cd5\u6765\u4ece\u89c2\u6d4b\u6570\u636e\u4e2d\u8fdb\u884c\u56e0\u679c\u7ed3\u6784\u4ee5\u53ca\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u4ed6\u4eec\u4ece\u56e0\u679c\u5173\u7cfb\u7684\u89d2\u5ea6\u8003\u8651\u4eba\u5de5\u667a\u80fd\u4e2d\u7684\u5404\u79cd\u5b66\u4e60\u95ee\u9898\uff0c\u5305\u62ec\u8fc1\u79fb\u5b66\u4e60\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u63a8\u8350\u7cfb\u7edf\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u5e76\u5e0c\u671b\u5bf9\u56e0\u679c\u5173\u7cfb\u7684\u8ba4\u77e5\u548c\u4f7f\u7528\u80fd\u628a\u4eba\u5de5\u667a\u80fd\u5e26\u5230\u4e00\u4e2a\u65b0\u7684\u9ad8\u5ea6\u3002<\/span><\/p>\n


<\/p>\n

\u672c\u6587\u4e3b\u8981\u5305\u542b\u56db\u4e2a\u90e8\u5206\u3002\u7b2c\u4e00\u90e8\u5206\u662f\u5bf9\u56e0\u679c\u53d1\u73b0\u4f20\u7edf\u5de5\u4f5c\u7684\u7b80\u4ecb\u5e76\u4e14\u5f15\u5165\u56e0\u679c\u8868\u5f81\uff1a\u4f55\u4e3a\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u5176\u4e0e\u56e0\u679c\u53d1\u73b0\u5173\u7cfb\u4f55\u5728\uff0c\u4ee5\u53ca\u4e3a\u4f55\u9700\u8981\u65b0\u65b9\u6cd5\u5b9e\u73b0\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u3002\u7b2c\u4e8c\u90e8\u5206\u7740\u91cd\u4e8e\u4ecb\u7ecd\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u7684\u6700\u65b0\u7814\u7a76\u8fdb\u5c55\u548c\u65b9\u6cd5\uff0c \u5305\u62ec\u5728\u72ec\u7acb\u540c\u5206\u5e03\u60c5\u5f62\u3001\u65f6\u95f4\u5e8f\u5217\u3001\u5206\u5e03\u8fc1\u79fb\u60c5\u5f62\u7684\u60c5\u51b5\u4e0b\u5982\u4f55\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002\u7b2c\u4e09\u90e8\u5206\u4ecb\u7ecd\u4e86\u5728\u8fc1\u79fb\u5b66\u4e60\u548c\u9002\u5e94\u6027\u5f3a\u5316\u5b66\u4e60\u4e0b\u7684\u5e94\u7528\u3002\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u4e3a\u673a\u5668\u5b66\u4e60\u63d0\u4f9b\u4e86\u66f4\u6709\u89e3\u91ca\u6027\u548c\u66f4\u6709\u6548\u7684\u65b9\u6cd5\u3002\u5728\u7b2c\u56db\u90e8\u5206\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86\u76ee\u524d\u7684\u5206\u6790\u5de5\u5177causal-learn\u4ee5\u53ca\u5bf9\u672a\u6765\u7814\u7a76\u5de5\u4f5c\u548c\u5e94\u7528\u573a\u666f\u7684\u5c55\u671b\u3002<\/span><\/p>\n


<\/p>\n

\u7531\u5468\u6653\u534e\u6559\u6388\u53d1\u8d77\u76842022\u6cdb\u592a\u5e73\u6d0b\u56e0\u679c\u63a8\u65ad\u5927\u4f1a<\/a>\uff0c\u5c06\u4e8e9\u670817\u65e5-9\u670818\u65e5\u5728\u7ebf\u4e3e\u529e\uff0c\u6c47\u805a\u56fd\u5185\u5916\u4ece\u4e8b\u56e0\u679c\u79d1\u5b66\u76f8\u5173\u5de5\u4f5c\u7684\u4e00\u7ebf\u7814\u7a76\u4eba\u5458\uff0c\u5171\u540c\u63a2\u8ba8\u56e0\u679c\u79d1\u5b66\u7684\u6700\u65b0\u8fdb\u5c55\u3002\u5c4a\u65f6\u96c6\u667a\u4ff1\u4e50\u90e8\u5c06\u5bf9\u4f1a\u8bae\u8fdb\u884c\u76f4\u64ad\u3002\u4f1a\u8bae\u4ecd\u5728\u5f00\u653e\u62a5\u540d\u4e2d\uff0c\u8bae\u7a0b\u53ca\u4f1a\u8bae\u8be6\u60c5\u89c1\u6587\u672b\u3002<\/strong><\/span><\/p>\n


<\/p>\n

\u7814\u7a76\u9886\u57df\uff1a\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u673a\u5668\u5b66\u4e60\uff0c\u56e0\u679c\u53d1\u73b0<\/span><\/strong><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/strong><\/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
<\/strong><\/section>\n<\/section>\n
\n

\u5f20\u5764\u56e2\u961f(\u9ec4\u78a7\u8587\u3001\u59da\u5dcd\u7136\u3001\u8c22\u5cf0\u3001\u90d1\u96e8\u5609\u3001\u5f20\u5764) <\/span><\/strong><\/span><\/strong>| \u4f5c\u8005<\/strong><\/span><\/strong><\/span><\/p>\n

\u9093\u4e00\u96ea<\/span> <\/span>| <\/strong><\/span>\u7f16\u8f91<\/strong><\/p>\n


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


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

\u76ee\u5f55<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

1. \u7b80<\/span>\u4ecb<\/span><\/strong><\/section>\n
2. \u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb<\/span><\/strong>
<\/span><\/section>\n
2.1 \u72ec\u7acb\u540c\u5206\u5e03\u60c5\u5f62<\/span><\/section>\n

(1) \u57fa\u4e8eLow-rank\u7684\u65b9\u6cd5<\/span><\/p>\n

(2) \u57fa\u4e8eGIN\u7684\u65b9\u6cd5<\/span><\/p>\n

(3) \u5c42\u7ea7\u7ed3\u6784\u5b66\u4e60\u7684\u65b9\u6cd5<\/span><\/p>\n

(4) \u5c0f\u7ed3<\/span><\/p>\n

2.2 \u65f6\u95f4\u5e8f\u5217<\/span><\/section>\n
(1) <\/span>\u4f20\u7edf\u56e0\u679c\u53d1\u73b0<\/span><\/section>\n
(2) <\/span>\u4ece\u9759\u6001\u65f6\u95f4\u5e8f\u5217\u4e2d\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb<\/span><\/section>\n
(3) <\/span>\u6570\u636e\u5206\u5e03\u53d8\u5316\u4e3a\u65f6\u95f4\u5e8f\u5217\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u5e26\u6765\u7684\u597d\u5904<\/span><\/section>\n
2.3 \u5206\u5e03\u8fc1\u79fb\u60c5\u5f62<\/span><\/section>\n

3. \u5e94\u7528<\/span><\/strong><\/p>\n

3.1 \u8fc1\u79fb\u5b66\u4e60<\/span><\/section>\n
3.2 \u9002\u5e94\u6027\u5f3a\u5316\u5b66\u4e60<\/span><\/section>\n

4. \u76ee\u524d\u7684\u5206\u6790\u5de5\u5177\u4ee5\u53ca\u5bf9\u672a\u6765\u7684\u5c55\u671b<\/span><\/strong><\/p>\n

4.1 Causal-learn<\/span><\/section>\n
(1) <\/span>\u5e73\u53f0\u4ecb\u7ecd<\/span><\/section>\n
(2) <\/span>\u7b80\u5355\u4e0a\u624b<\/span><\/section>\n
4.2 <\/span>\u76ee\u524d\u7684\u96be\u70b9\u548c\u5c06\u6765\u7684\u91cd\u70b9<\/span><\/section>\n


<\/p>\n


<\/p>\n

\n
\n
\n
\n
\n

<\/section>\n

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

1. \u7b80\u4ecb<\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

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

<\/section>\n

\u56e0\u679c\u8868\u5f81\u5b66\u4e60<\/span>\uff08causal representation learning\uff09<\/span>\u662f\u8fde\u63a5\u56e0\u679c\u79d1\u5b66\u4e0e\u6df1\u5ea6\u5b66\u4e60\u7684\u6865\u6881\u3002\u8fd1\u5341\u5e74\u6765\uff0c\u6df1\u5ea6\u5b66\u4e60<\/span>\uff08deep learning\uff09<\/span>\u6210\u4e3a\u4e86\u8ba1\u7b97\u673a\u89c6\u89c9<\/span>\uff08CV\uff09<\/span>\u3001\u81ea\u7136\u8bed\u8a00\u5904\u7406<\/span>\uff08NLP\uff09<\/span>\u7b49\u4eba\u5de5\u667a\u80fd<\/span>\uff08AI\uff09<\/span>\u57fa\u7840\u9886\u57df\u7684\u6838\u5fc3\u6280\u672f\uff0c\u540c\u65f6\u4e3a\u673a\u5668\u4eba\uff0c\u65e0\u4eba\u9a7e\u9a76\uff0c\u4eba\u673a\u4ea4\u4e92\u7cfb\u7edf\uff0c\u865a\u62df\u73b0\u5b9e\u3001\u751f\u7269\u5236\u836f\u3001\u667a\u6167\u57ce\u5e02\u7b49\u884c\u4e1a\u5e94\u7528\u8fdb\u884c\u667a\u6167\u8d4b\u80fd\u3002\u6df1\u5ea6\u5b66\u4e60\u662f\u4e00\u79cd\u4f7f\u7528\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u901a\u8fc7\u68af\u5ea6\u53cd\u4f20\u7684\u4fe1\u53f7\u53cd\u9988\u5c06\u975e\u7ed3\u6784\u5316\u6570\u636e\uff0c\u5982\u56fe\u7247\u3001\u8bed\u8a00\u6587\u5b57\u7b49\uff0c\u8f6c\u5316\u4e3a\u53ef\u88ab\u673a\u5668\u5904\u7406\u7684\u8868\u5f81<\/span>\uff08representation\uff09<\/span>\uff0c\u5e76\u4f7f\u7528\u8fd9\u4e9b\u8868\u5f81\u89e3\u51b3\u8bc6\u522b\u3001\u7406\u89e3\u7b49\u4e0b\u6e38\u4efb\u52a1\u7684\u8ba1\u7b97\u65b9\u6cd5\u30022015\u5e74\uff0c\u5728\u7531Geoffrey Hinton, Yann LeCun and Yoshua Bengio<\/span>\uff08\u56fe\u7075\u5956\u83b7\u5f97\u8005\uff0c\u6df1\u5ea6\u5b66\u4e60\u8457\u540d\u4eba\u7269\uff09<\/span>\u5171\u540c\u7f72\u540d\u7684Nature<\/em>\u8bba\u6587\u300aDeep Learning\u300b\u4e2d\uff0c\u8868\u5f81\u5b66\u4e60<\/span> (representation learning\uff09<\/span>\u88ab\u5f52\u4e3a\u6df1\u5ea6\u5b66\u4e60\u548c\u73b0\u4ee3AI\u6280\u672f\u53d6\u5f97\u5de8\u5927\u6210\u529f\u7684\u4e3b\u8981\u539f\u56e0\u3002\u8868\u5f81\u5b66\u4e60\u7684\u76ee\u6807\u662f\u5c06\u9ad8\u7ef4\u7684\u539f\u59cb\u6570\u636e\u964d\u7ef4\u6210\u4f4e\u7ef4\u7279\u5f81\uff0c\u5728\u4fdd\u7559\u4fe1\u606f\u7684\u540c\u65f6\u8fc7\u6ee4\u6389\u539f\u59cb\u6570\u636e\u4e2d\u7684\u566a\u58f0\u3002<\/span><\/p>\n


<\/p>\n

\u7136\u800c\uff0c\u901a\u8fc7\u6269\u5927\u6570\u636e\u548c\u6a21\u578b\u7684\u89c4\u6a21\u5b66\u5f97\u7684\u8868\u5f81\u5f80\u5f80\u53ea\u80fd\u5bf9\u6570\u636e\u8fdb\u884c\u7f16\u7801\u3001\u538b\u7f29\u6216\u8bb0\u5fc6\u3002\u8fd9\u4e9b\u8868\u5f81\u4e0d\u5177\u6709\u53ef\u4f9b\u673a\u5668\u8fdb\u884c\u903b\u8f91\u63a8\u7406\u548c\u89c4\u5212\u7684\u9ad8\u9636\u8bed\u4e49\u56e0\u5b50<\/span>\uff08semantic concept\uff09<\/span>\uff0c\u4e5f\u4e0d\u5177\u5907\u50cf\u4eba\u7c7b\u4e00\u6837\u5728\u65b0\u73af\u5883\u4e0b\u7075\u6d3b\u89e3\u51b3\u95ee\u9898\u7684\u53d8\u901a\u80fd\u529b\u548c\u9c81\u68d2\u6027\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u4f7f\u6df1\u5ea6\u5b66\u4e60\u80fd\u50cf\u4eba\u7c7b\u4e00\u6837\u8fdb\u884c\u6709\u610f\u8bc6\u7684\u63a8\u7406\u3001\u601d\u8003\u4e0e\u5224\u65ad\u662f\u4e0b\u4e00\u4ee3AI\u6025\u9700\u89e3\u51b3\u7684\u95ee\u9898\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u56e0\u679c<\/span>\uff08causal\uff09<\/span>\u6a21\u578b\uff0c\u63d0\u4f9b\u4e86\u4e00\u5957\u7cfb\u7edf\u6027\u7684\u3001\u57fa\u4e8e\u7edf\u8ba1\u7684\u56e0\u679c\u63a8\u7406\u548c\u601d\u8003\u7684\u8ba1\u7b97\u65b9\u6cd5\u3002\u7136\u800c\uff0c\u56e0\u679c\u6a21\u578b\u5f80\u5f80\u53ea\u80fd\u5904\u7406\u7ed3\u6784\u5316\u7684\u6570\u636e\uff0c\u5e76\u4e0d\u80fd\u5904\u7406\u751f\u6d3b\u4e2d\u5e38\u89c1\u7684\u9ad8\u7ef4\u7684\u539f\u59cb\u6570\u636e\uff0c\u6bd4\u5982\u56fe\u50cf\u3002\u4e8e\u662f\uff0c\u5c06\u8868\u5f81\u5b66\u4e60\u548c\u56e0\u679c\u6a21\u578b\u8fdb\u884c\u878d\u5408\uff0c\u5c06\u56fe\u50cf\u8fd9\u6837\u7684\u539f\u59cb\u6570\u636e\u8f6c\u5316\u4e3a\u53ef\u7528\u4e8e\u56e0\u679c\u6a21\u578b\u7684\u7ed3\u6784\u5316\u53d8\u91cf\uff0c\u8d4b\u4e88AI\u5982\u540c\u4eba\u7c7b\u4e00\u6837\u6709\u610f\u8bc6\u7684\u63a8\u7406\u548c\u601d\u8003\u7684\u80fd\u529b\uff0c\u6210\u4e3a\u4e86\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u8fd9\u4e00\u65b0\u5174\u5b66\u79d1\u7684\u4e3b\u8981\u76ee\u6807\u3002\u5982\u679c\u6211\u4eec\u80fd\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5f88\u597d\u5730\u5c06\u56e0\u679c\u63a8\u65ad\u4e0e\u673a\u5668\u5b66\u4e60\u7ed3\u5408\u8d77\u6765\uff0c\u6784\u5efa\u4e0b\u4e00\u4ee3\u7075\u6d3b\uff0c\u53ef\u4fe1\u7684AI\u3002<\/span><\/p>\n


<\/p>\n

\u5728\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u4e2d\uff0c\u6211\u4eec\u901a\u5e38\u5047\u8bbe\u6570\u636e\u662f\u7531\u56e0\u679c\u76f8\u5173\u7684\u3001\u6ee1\u8db3\u4e00\u5b9a\u6761\u4ef6\u7684\u7ed3\u6784\u56e0\u679c\u6a21\u578b<\/span>\uff08Structural Causal Model\uff0cSCM\uff09<\/span>\u7684\u56e0\u679c\u9690\u53d8\u91cf\uff0c\u901a\u8fc7\u975e\u7ebf\u6027\u7684\u6620\u5c04\u6765\u4ea7\u751f\u3002\u5982\u679c\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684SCM\u80fd\u4ece\u539f\u59cb\u6570\u636e\u4e2d\u5b66\u4e60\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u4f30\u8ba1\u5728\u5e72\u9884\u4e86\u8fd9\u4e9b\u53d8\u91cf\u540e\u7684\u6570\u636e\u5206\u5e03\uff0c\u6216\u63a8\u65ad\u6307\u5b9a\u6570\u636e\u70b9\u7684\u53cd\u4e8b\u5b9e\u7ed3\u679c\uff0c\u6bd4\u5982\u91cd\u65b0\u7ec4\u5408\u56e0\u679c\u9690\u53d8\u91cf\u751f\u6210\u4e8b\u5b9e\u4e2d\u4e0d\u5b58\u5728\u7684\u6570\u636e\uff0c\u6216\u8005\u56de\u7b54Why\uff0cWhat if\u7b49\u9700\u8981\u663e\u5f0f\u63a8\u7406\u7684\u95ee\u9898\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u4e0e\u56e0\u679c\u53d1\u73b0<\/span>\uff08Causal Discovery\uff09<\/span>\u6709\u5bc6\u4e0d\u53ef\u5206\u7684\u5173\u7cfb\u3002\u4e00\u65b9\u9762\uff0c\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u662f\u5728\u6709\u5f88\u591a\u6df7\u6dc6\u56e0\u5b50<\/span>\uff08confounder\uff09<\/span>\u4e0b\u7684\u56e0\u679c\u53d1\u73b0\u95ee\u9898\u7684\u4e00\u4e2a\u7279\u4f8b\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5728\u6ca1\u6709\u9886\u57df\u77e5\u8bc6\u7684\u60c5\u51b5\u4e0b\uff0c\u5b66\u4e60\u6574\u4e2a\u9690\u53d8\u91cf\u7a7a\u95f4\u4e2d\u7684\u7ed3\u6784\u56e0\u679c\u6a21\u578b\u662f\u6781\u5176\u56f0\u96be\u7684\u3002\u56e0\u6b64\uff0c\u56e0\u679c\u53d1\u73b0\u4e2d\u5e38\u7528\u7684\u5047\u8bbe\uff0c\u6bd4\u5982Sparsity\u3001Minimal Change Principle\u548cIndependent Causal Mechansim\uff0c\u901a\u5e38\u53ef\u4ee5\u6210\u4e3a\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u7684\u9886\u57df\u77e5\u8bc6\u548c\u4f7f\u7528\u7684\u5f52\u7eb3\u504f\u7f6e<\/span>\uff08Inductive Bias\uff09[Sch\u00f6lkopf et al., 2021]<\/span>\u3002<\/span><\/p>\n


<\/p>\n

\u56e0\u679c\u53d1\u73b0\u4f20\u7edf\u5de5\u4f5c\u7b80\u4ecb<\/strong><\/span><\/p>\n


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

\u5bfb\u627e\u56e0\u679c\u5173\u7cfb\u7684\u4f20\u7edf\u65b9\u6cd5\u662f\u901a\u8fc7\u968f\u673a\u5bf9\u7167\u5b9e\u9a8c\u3002\u7136\u800c\uff0c\u968f\u673a\u5bf9\u7167\u5b9e\u9a8c\u901a\u5e38\u9700\u8981\u8017\u8d39\u5927\u91cf\u7684\u65f6\u95f4\u548c\u8d44\u6e90\uff0c\u6709\u65f6\u751a\u81f3\u53ef\u80fd\u6d89\u53ca\u4f26\u7406\u95ee\u9898\u3002\u4f8b\u5982\uff0c\u57fa\u56e0\u6572\u5165\u662f\u4e00\u79cd\u5e7f\u6cdb\u7528\u4e8e\u521b\u5efa\u75be\u75c5\u6a21\u578b\u7684\u6280\u672f\u3002\u901a\u5e38\u5bfc\u81f4\u75be\u75c5\u7684\u4e0d\u662f\u5355\u4e2a\u57fa\u56e0\uff0c\u800c\u662f\u7531\u591a\u4e2a\u57fa\u56e0\u7ec4\u6210\u7684\u7f51\u7edc\u3002\u8981\u627e\u5230\u8fd9\u4e2a\u7f51\u7edc\u53ef\u80fd\u9700\u8981\u592a\u591a\u7684\u5b9e\u9a8c\u624d\u80fd\u53ef\u884c\u3002\u56e0\u6b64\uff0c\u5fc5\u987b\u5bfb\u6c42\u66ff\u4ee3\u65b9\u6cd5\u2014\u2014\u4ece\u89c2\u6d4b\u6570\u636e\u4e2d\u53d1\u73b0\u56e0\u679c\u5173\u7cfb\uff0c\u79f0\u4e3a\u56e0\u679c\u53d1\u73b0<\/span>\uff08causal discovery\uff09<\/span>\uff0c\u56e0\u4e3a\u89c2\u6d4b\u6570\u636e\u66f4\u5bb9\u6613\u83b7\u5f97\u3002<\/span><\/p>\n


<\/p>\n

\u7ecf\u5178\u7684\u56e0\u679c\u53d1\u73b0\u65b9\u6cd5\u901a\u5e38\u662f\u5bfb\u627e\u89c2\u6d4b\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u5176\u5927\u81f4\u5206\u4e3a\u4e24\u7c7b\u3002\u5728 1980 \u5e74\u4ee3\u672b\u548c 1990 \u5e74\u4ee3\u521d\uff0c\u4eba\u4eec\u6ce8\u610f\u5230\uff0c\u5728\u9002\u5f53\u7684\u5047\u8bbe\u4e0b\uff0c\u53ef\u4ee5\u6839\u636e\u53d8\u91cf\u4e4b\u95f4\u7684\u6761\u4ef6\u72ec\u7acb\u6027\u5173\u7cfb\u6062\u590d\u6f5c\u5728\u56e0\u679c\u7ed3\u6784\u7684\u9a6c\u5c14\u53ef\u592b\u7b49\u4ef7\u7c7b [Spirtes et al., 1993]\u3002\u8fd9\u4ea7\u751f\u4e86\u57fa\u4e8e\u7ea6\u675f\u7684\u65b9\u6cd5<\/span>\uff08constraint-based method\uff09<\/span>\uff0c\u8be5\u65b9\u6cd5\u5229\u7528\u6761\u4ef6\u72ec\u7acb\u6027\u6d4b\u8bd5<\/span>\uff08conditional independence tests\uff09\u548c\u79bb\u6563\u641c\u7d22\uff08discrete search\uff09<\/span><\/span>\u6765\u8fdb\u884c\u56e0\u679c\u53d1\u73b0\u3002\u7531\u6b64\u4ea7\u751f\u7684\u7b49\u4ef7\u7c7b\u53ef\u80fd\u5305\u542b\u591a\u4e2a \u6709\u5411\u65e0\u73af\u56fe<\/span>\uff08DAG\uff0c\u6216\u5176\u4ed6\u8868\u793a\u56e0\u679c\u7ed3\u6784\u7684\u56fe\u5bf9\u8c61\uff09<\/span>\uff0c\u8fd9\u4e9bDAG\u5171\u4eab\u76f8\u540c\u7684\u6761\u4ef6\u72ec\u7acb\u6027\u5173\u7cfb\u3002\u6240\u9700\u7684\u5047\u8bbe\u5305\u62ec\u56e0\u679c\u9a6c\u5c14\u53ef\u592b\u6761\u4ef6<\/span>\uff08causal Markov condition\uff09<\/span>\u548c\u5fe0\u5b9e\u5ea6\u5047\u8bbe<\/span>\uff08faithfulness assumption\uff09<\/span>\uff0c\u8fd9\u4e24\u4e2a\u5047\u8bbe\u5efa\u7acb\u4e86\u56e0\u679c\u56fe\u4e2d\u7684d-\u5206\u79bb\u5c5e\u6027\u4e0e\u6570\u636e\u4e2d\u7684\u7edf\u8ba1\u72ec\u7acb\u6027\u5c5e\u6027\u4e4b\u95f4\u7684\u5bf9\u5e94\u5173\u7cfb\u3002\u76f8\u53cd\uff0c\u57fa\u4e8e\u5206\u6570\u7684\u65b9\u6cd5 [Chickering, 2003, Heckerman et al., 1995] \u4e0d\u662f\u4f7f\u7528\u7edf\u8ba1\u6d4b\u8bd5\uff0c\u800c\u662f\u641c\u7d22\u5728\u67d0\u4e9b\u8bc4\u5206\u6807\u51c6\u4e0b\u7ed9\u51fa\u6700\u9ad8\u5206\u6570\u7684\u7b49\u4ef7\u7c7b\uff0c\u4f8b\u5982\u91c7\u7528\u8d1d\u53f6\u65af\u4fe1\u606f\u6807\u51c6 [BIC, Schwartz, 1978]\u3001\u7ed9\u5b9a\u6570\u636e\u7684\u56fe\u7684\u540e\u9a8c [Heckerman et al., 1997] \u548c\u5e7f\u4e49\u8bc4\u5206\u51fd\u6570 [generalized score, Huang et al., 2018]\u3002<\/span><\/p>\n


<\/p>\n

\u53e6\u4e00\u7ec4\u65b9\u6cd5\u662f\u57fa\u4e8e\u51fd\u6570\u56e0\u679c\u6a21\u578b<\/span>\uff08functional causal model, FCM\uff09<\/span>\uff0c\u5c06\u6548\u679c\u8868\u793a\u4e3a\u76f4\u63a5\u539f\u56e0<\/span> (direct cause) <\/span>\u7684\u51fd\u6570\u4ee5\u53ca\u72ec\u7acb\u7684\u566a\u58f0\u9879\u3002\u7814\u7a76\u8868\u660e\uff0c\u901a\u8fc7\u9002\u5f53\u5730\u7ea6\u675f\u6a21\u578b\u7c7b\uff0c\u56e0\u679c\u65b9\u5411\u662f\u53ef\u8bc6\u522b\u7684\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5f53\u5728\u6b63\u786e\u7684\u56e0\u679c\u65b9\u5411\u4e0a\u4f30\u8ba1 FCM \u65f6\uff0c\u4f30\u8ba1\u7684\u566a\u58f0\u9879\u4e0e\u5047\u8bbe\u539f\u56e0\u72ec\u7acb\uff0c\u4f46\u5728\u60f3\u53cd\u7684\u65b9\u5411\u4e0a\u4e0d\u6210\u7acb\u3002\u53ef\u8bc6\u522b\u7684\u56e0\u679c\u6a21\u578b\u5305\u62ec\u7ebf\u6027\u975e\u9ad8\u65af\u65e0\u73af\u6a21\u578b [LiNGAM, Shimizu et al., 2006]\u3001\u975e\u7ebf\u6027\u52a0\u6027\u566a\u58f0\u6a21\u578b [additive noise causal model, Hoyer et al., 2009, Zhang and Hyv\u00e4rinen, 2009a] \u548c\u540e\u975e\u7ebf\u6027\u6a21\u578b [post-nonlinear causal model, Zhang and Chan, 2006, Zhang and Hyv\u00e4rinen, 2009b]\u3002\u4f46\u662f\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u5982\u679c\u51fd\u6570\u56e0\u679c\u6a21\u578b\u7684\u51fd\u6570\u7a7a\u95f4\u6ca1\u6709\u7ea6\u675f\uff0c\u5219\u65e0\u6cd5\u8bc6\u522b\u56e0\u679c\u65b9\u5411\uff0c\u56e0\u4e3a\u603b\u662f\u53ef\u4ee5\u5728\u53cd\u5411\u4e0a\u627e\u5230\u72ec\u7acb\u4e8e\u9884\u6d4b\u53d8\u91cf\u7684\u566a\u58f0\u9879 [Zhang et al, 2015]\u3002<\/span><\/p>\n


<\/p>\n

\u4e0a\u9762\u4ecb\u7ecd\u7684\u65b9\u6cd5\u5df2\u88ab\u6269\u5c55\u5230\u66f4\u4e00\u822c\u7684\u573a\u666f\u3002\u6bd4\u5982\u8bf4LiNGAM\u5df2\u88ab\u6269\u5c55\u5230\u6709\u73af\u7684\u56e0\u679c\u56fe [Lacerda et al., 2008] \u548c\u5b58\u5728\u6f5c\u5728\u6df7\u6742\u56e0\u7d20\u7684\u60c5\u51b5 [Hoyer et al., 2008] \u3002\u57fa\u4e8eLiNGAM\u62d3\u5c55\u7684Specific and Shared Causal Relation Modeling [SSCM, Huang et al., 2019] \u4e0d\u4ec5\u53ef\u4ee5\u63d0\u4f9b\u6574\u4f53\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u8fd8\u53ef\u4ee5\u63d0\u4f9b\u9488\u5bf9\u4e2a\u4f53\u7684\u4e2a\u6027\u5316\u7684\u56e0\u679c\u77e5\u8bc6\uff0c \u4ee5\u53ca\u57fa\u4e8e\u56e0\u679c\u5173\u7cfb\u6765\u505a\u805a\u7c7b\u3002\u540c\u65f6\uff0c\u7814\u7a76\u8868\u660e\u5373\u4f7f\u5b58\u5728\u9009\u62e9\u504f\u5dee\u7684\u60c5\u51b5\u4e0b\uff0c\u56e0\u679c\u6a21\u578b<\/span>\uff08\u5305\u62ec\u56e0\u679c\u65b9\u5411\uff09<\/span>\u4e5f\u662f\u53ef\u80fd\u8bc6\u522b\u7684 [Zhang et al., 2016]\u3002<\/span><\/p>\n


<\/p>\n

\u76ee\u524d\u7684\u56e0\u679c\u53d1\u73b0\u65b9\u6cd5\u96c6\u4e2d\u5728\u5bfb\u627e\u89c2\u6d4b\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c \u4f46\u662f\u5728\u73b0\u5b9e\u4e16\u754c\u7684\u95ee\u9898\u4e2d\uff0c\u5f88\u591a\u76f8\u5173\u7684\u7279\u5f81\u53ef\u80fd\u6ca1\u6709\u88ab\u89c2\u5bdf\u5230\uff0c\u5e76\u4e14\u4e00\u4e9b\u89c2\u6d4b\u53d8\u91cf\u53ef\u80fd\u4e0d\u662f\u6f5c\u5728\u7684\u56e0\u679c\u53d8\u91cf\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u4e0d\u80fd\u76f4\u63a5\u5c06\u56fe\u50cf\u50cf\u7d20\u89c6\u4e3a\u56e0\u679c\u53d8\u91cf\u3002\u56e0\u6b64\uff0c\u6b63\u5982\u5728\u7b2c\u4e00\u90e8\u5206\u4e2d\u6240\u8bf4\u7684\uff0c\u6211\u4eec\u60f3\u4ece\u6d4b\u91cf\u7684\u9ad8\u7ef4\u53d8\u91cf\u4e2d\u5b66\u4e60\u6f5c\u5728\u7684\u56e0\u679c\u8868\u5f81<\/span>\uff08causal representation\uff09<\/span>\uff0c\u4ee5\u53ca\u5b83\u4eec\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u5b83\u5728\u901a\u7528\u4eba\u5de5\u667a\u80fd\u548c\u79d1\u5b66\u9886\u57df\u90fd\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\u3002\u4f8b\u5982\uff0c\u5728 AI \u4e2d\uff0c\u6211\u4eec\u5e0c\u671b\u4ece\u9ad8\u7ef4\u89c6\u9891\u5e8f\u5217\u4e2d\u81ea\u52a8\u63d0\u53d6\u5e95\u5c42\u7684\u4f4e\u7ef4\u56e0\u679c\u53d8\u91cf\u6216\u6982\u5ff5\uff0c\u8fd9\u4e9b\u53d8\u91cf\u6216\u6982\u5ff5\u5bf9\u4e8e\u89c6\u9891\u7406\u89e3\u81f3\u5173\u91cd\u8981\uff0c\u4ece\u800c\u4fc3\u8fdb\u4e0b\u6e38\u9884\u6d4b\u6216\u51b3\u7b56\u4efb\u52a1\u3002\u5728\u795e\u7ecf\u79d1\u5b66\u4e2d\uff0c\u4ece fMRI \u8bb0\u5f55\u4e2d\u6d4b\u91cf\u51fa\u6570\u4ee5\u4e07\u8ba1\u7684\u4f53\u7d20\uff0c\u4e00\u4e2a\u5173\u952e\u95ee\u9898\u662f\u5982\u4f55\u8bc6\u522b\u548c\u5206\u5c42\u805a\u7c7b\u6f5c\u5728\u7684\u5927\u8111\u529f\u80fd\u533a\u57df\u5e76\u53d1\u73b0\u4fe1\u606f\u6d41\u3002<\/span><\/p>\n


<\/p>\n

\u5728\u672c\u7bc7\u63a8\u9001\u4e2d\uff0c\u6211\u4eec\u5c06\u7ed3\u5408\u4f20\u7edf\u56e0\u679c\u53d1\u73b0\uff0c\u4ece\u65b9\u6cd5\uff0c\u5e94\u7528\u548c\u5de5\u5177\u89d2\u5ea6\uff0c\u8be6\u7ec6\u63a2\u8ba8\u5982\u4f55\u5728\u72ec\u7acb\u540c\u5206\u5e03\u60c5\u5f62\u3001\u65f6\u95f4\u5e8f\u5217\u548c\u5206\u5e03\u8fc1\u79fb\u60c5\u5f62\u4e0b\u8fdb\u884c\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u6211\u4eec\u603b\u7ed3\u4e86\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u5728\u8fc1\u79fb\u5b66\u4e60\uff0c\u9002\u5e94\u6027\u5f3a\u5316\u5b66\u4e60\u7684\u5e94\u7528\u3002\u6700\u540e\uff0c\u6211\u4eec\u5bf9\u76ee\u524d\u7684\u53ef\u76f4\u63a5\u4f7f\u7528\u7684\u5206\u6790\u5de5\u5177\u505a\u7b80\u8981\u4ecb\u7ecd\u548c\u5e76\u5bf9\u672a\u6765\u7684\u7814\u7a76\u548c\u5e94\u7528\u65b9\u5411\u8fdb\u884c\u5c55\u671b\u3002<\/span><\/p>\n


<\/p>\n

\n
\n
\n

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

\n
\n
\n
\n

LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep learning.” nature 521.7553 (2015): 436-444.<\/span><\/p>\n

Sch\u00f6lkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., & Bengio, Y. (2021). Toward causal representation learning. Proceedings of the IEEE, 109(5), 612-634.<\/span><\/p>\n

P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction, and Search. Spring-Verlag Lectures in Statistics, 1993.<\/span><\/p>\n

D. M. Chickering. Optimal structure identification with greedy search. Journal of Machine Learning Research, 3:507\u2013554, 2003.<\/span><\/p>\n

D. Heckerman, D. Geiger, and D. M. Chickering. Learning bayesian networks: The combination of knowledge and statistical data. Machine Learning, 20:197\u2013243, 1995.<\/span><\/p>\n

G. Schwartz. Estimating the dimension of a model. The Annals of Statistics, 5:461\u2013464, 1978.<\/span><\/p>\n

D. Heckerman, C. Meek, and G. Cooper. Interventions and causal inference. In Innovations in Machine Learning, pages 1\u201328, 1997.<\/span><\/p>\n

B. Huang, K. Zhang, Y. Lin, B. Scho\u0308lkopf, and C. Glymour. Generalized score functions for causal discovery. In KDD, pages 1551\u20131560, 2018.<\/span><\/p>\n

S. Shimizu, P.O. Hoyer, A. Hyva\u0308rinen, and A.J. Kerminen. A linear non-Gaussian acyclic model for causal discovery. Journal of Machine Learning Research, 7:2003\u20132030, 2006.<\/span><\/p>\n

P. O. Hoyer, D. Janzing, J. Mooji, J. Peters, and B. Scho\u0308lkopf. Nonlinear causal discovery with additive noise models. In Advances in Neural Information Processing Systems 21, Vancouver, B.C., Canada, 2009.<\/span><\/p>\n

K. Zhang and A. Hyva\u0308rinen. Acyclic causality discovery with additive noise: An information- theoretical perspective. In Proc. European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD) 2009, Bled, Slovenia, 2009a.<\/span><\/p>\n

K. Zhang and L. Chan. Extensions of ICA for causality discovery in the hong kong stock market. In Proc. 13th International Conference on Neural Information Processing (ICONIP 2006), 2006.<\/span><\/p>\n

K. Zhang and A. Hyva\u0308rinen. On the identifiability of the post-nonlinear causal model. In Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence, Montreal, Canada, 2009b.<\/span><\/p>\n

K. Zhang, Z. Wang, J. Zhang and B. Sch\u00f6lkopf. On estimation of functional causal models: general results and application to the post-nonlinear causal model. ACM Transactions on Intelligent Systems and Technology (TIST), 7(2), 1-22.<\/span><\/p>\n

G. Lacerda, P. Spirtes, J. Ramsey, and P. O. Hoyer. Discovering cyclic causal models by independent components analysis. In Proceedings of the 24th Conference on Uncertainty in Artificial Intelligence (UAI2008), Helsinki, Finland, 2008.<\/span><\/p>\n

P. O. Hoyer, S. Shimizu, A. J. Kerminen, and M. Palviainen. Estimation of causal effects using linear non-gaussian causal models with hidden variables. International Journal of Approximate Reasoning, 49:362\u2013378, 2008.<\/span><\/p>\n

B. Huang, K. Zhang, P. Xie, M. Gong, E. Xing, and C. Glymour. Specific and shared causal relation modeling and mechanism-based clustering. In Advances in Neural Information Processing Systems (NeurIPS), 2019.<\/span><\/p>\n

K. Zhang, J. Zhang, B. Huang, B. Scho\u0308lkopf, and C. Glymour. On the identifiability and estimation of functional causal models in the presence of outcome-dependent selection. UAI, 2016.<\/span><\/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


<\/p>\n

\n
\n
\n
\n
\n

<\/section>\n

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

2. \u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb<\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

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


<\/p>\n

\u6309\u6240\u4f7f\u7528\u6570\u636e\u7684\u4e0d\u540c\u6027\u8d28\uff0c\u4e0b\u9762\u6211\u4eec\u5c06\u628a\u8fd1\u51e0\u5e74\u63d0\u51fa\u7684\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u7684\u65b9\u6cd5\u5206\u6210\u4e09\u7c7b\u8fdb\u884c\u8ba8\u8bba\uff0c\u5b83\u4eec\u5206\u522b\u9002\u7528\u4e8e\u72ec\u7acb\u540c\u5206\u5e03\u6570\u636e\u3001\u65f6\u95f4\u5e8f\u5217\uff0c\u4ee5\u53ca\u6709\u5206\u5e03\u8fc1\u79fb\u6027\u8d28\u7684\u6570\u636e<\/span>\uff08\u6bd4\u5982\u975e\u5e73\u7a33\u6216\u5f02\u6784\u6570\u636e\uff09<\/span>\u3002<\/span><\/section>\n

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

2.1 \u72ec\u7acb\u540c\u5206\u5e03\u60c5\u5f62<\/span><\/strong><\/span><\/h2>\n

<\/span><\/section>\n
\u89e3\u51b3\u542b\u6709\u9690\u53d8\u91cf<\/span> (\u9690\u6df7\u6dc6\u56e0\u5b50) <\/span>\u7684\u4e00\u4e2a\u7ecf\u5178\u65b9\u5f0f\u662f\u57fa\u4e8e\u6761\u4ef6\u72ec\u7acb\u6027\u6d4b\u8bd5\uff0c\u5982FCI <\/span>(Fast Causal Inference) <\/span>\u7b97\u6cd5[Spirtes et al., 1995]\u3002\u8be5\u65b9\u6cd5\u9996\u5148\u501f\u52a9d-\u5206\u79bb\u51c6\u5219\u5b66\u4e60\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u9aa8\u67b6\u56fe\uff0c\u7136\u540e\u901a\u8fc7 V-\u7ed3\u6784\u548c\u4e00\u4e9b\u5b9a\u4f4d\u51c6\u5219\u8fdb\u4e00\u6b65\u63a8\u65ad\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u65b9\u5411\uff0c\u6700\u7ec8\u5f97\u5230\u56e0\u679c\u7956\u5148\u56fe\u3002\u5c3d\u7ba1\u8fd9\u7c7b\u65b9\u6cd5\u5728\u8bb8\u591a\u9886\u57df\u7684\u5230\u4e86\u5e94\u7528\uff0c\u7136\u800c\u8fd9\u7c7b\u65b9\u6cd5\u7684\u8f93\u51fa\u662f\u89c2\u6d4b\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u800c\u5ffd\u7565\u4e86\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002\u8fd1\u671f\uff0c\u7814\u7a76\u5b66\u8005\u5229\u7528\u9002\u7528\u4e8e\u9690\u53d8\u91cf\u4e0b\u65b0\u7684\u201cd-\u5206\u79bb\u51c6\u5219\u201d\uff0c\u5305\u62ecLow-rank\u6761\u4ef6\uff0cGeneralized Independent Noise  <\/span>(GIN) <\/span>\u6761\u4ef6\u7b49\uff0c\u7ed9\u51fa\u4e86\u7ebf\u6027\u7cfb\u7edf\u4e0b\u9690\u53d8\u91cf\u95f4\u56e0\u679c\u7ed3\u6784\u7684\u8bc6\u522b\u6027\u6761\u4ef6\u548c\u5b66\u4e60\u65b9\u6cd5\uff0c\u63a5\u4e0b\u6765\u4f1a\u5728\u672c\u8282\u9010\u4e00\u4ecb\u7ecd\u3002<\/span><\/section>\n

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

(1) \u57fa\u4e8eLow-rank\u7684\u65b9\u6cd5<\/span><\/strong><\/h3>\n

<\/span><\/section>\n
\u56e0\u5b50\u5206\u6790\u662f\u89e3\u51b3\u9690\u53d8\u91cf\u95f4\u56e0\u679c\u7ed3\u6784\u7684\u7ecf\u5178\u65b9\u6cd5\uff0c\u7136\u800c\u8be5\u65b9\u6cd5\u8f93\u51fa\u7684\u9690\u53d8\u91cf\u7ed3\u6784\u4e0d\u80fd\u591f\u4fdd\u8bc1\u662f\u80cc\u540e\u771f\u5b9e\u53d1\u751f\u56e0\u679c\u673a\u5236\u7684\u7ed3\u6784[Spirtes et al., 2000; Silva et al., 2006]\u3002Silva\u7b49\u4eba[Silva et al., 2006]\u5f00\u521b\u6027\u5730\u5229\u7528Tetrad\u6761\u4ef6[Spearman,1928]\uff0c\u63d0\u51fa\u4e86\u4e00\u79cd\u4e24\u9636\u6bb5\u5b66\u4e60\u9690\u53d8\u91cf\u7ed3\u6784\u7684\u6846\u67b6\uff0c\u5373BPC<\/span>\uff08BulidPureClusters\uff09<\/span>\u7b97\u6cd5+MIMBulid\u7b97\u6cd5\u3002\u5177\u4f53\u6765\u8bb2\uff1a\u4ee5\u56fe1\u4e2d\u4e0a\u7684\u5b50\u56fe\u4e3a\u4f8b\u3002\u7b2c\u4e00\u9636\u6bb5\u5229\u7528Tetrad\u6761\u4ef6\u8bbe\u8ba1\u4e86\u4e09\u79cd\u5224\u65ad\u6d4b\u91cf\u53d8\u91cf\u95f4\u662f\u5426\u5171\u4eab\u540c\u4e00\u4e2a\u7236\u8282\u70b9\u7684\u89c4\u5219\u4ece\u800c\u53bb\u5b66\u4e60\u7eaf\u7684\u6d4b\u91cf\u6a21\u578b<\/span>\uff08Pure Measurement Model\uff0c\u5305\u542b\u9690\u53d8\u91cf\u4ee5\u53ca\u9690\u53d8\u91cf\u4e0e\u6d4b\u91cf\u53d8\u91cf\u4e4b\u95f4\u7684\u5173\u7cfb\uff09<\/span>\uff0c\u5373\u5b50\u56fe (a)\uff1b\u7b2c\u4e8c\u9636\u6bb5\u628a\u6bcf\u4e2a\u9690\u53d8\u91cf\u6240\u5bf9\u5e94\u7684\u6d4b\u91cf\u53d8\u91cf\u4e3a\u4ee3\u7406\u53d8\u91cf\uff0c\u8fdb\u4e00\u6b65\u5b66\u4e60\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u7ed3\u6784 <\/span>\uff08Structural Model\uff0c\u5305\u542b\u9690\u53d8\u91cf\u53ca\u5176\u56e0\u679c\u5173\u7cfb\uff09<\/span>\uff0c\u5373\u5b50\u56fe (b)\u3002<\/span><\/section>\n

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

\u56fe1\uff1a\u4e24\u9636\u6bb5\u5b66\u4e60\u9690\u53d8\u91cf\u95f4\u56e0\u679c\u7ed3\u6784\u793a\u4f8b\u56fe\u3002\u6700\u4e0a\u9762\u4e3a\u771f\u5b9e\u56e0\u679c\u56fe\uff1b\u5b50\u56fe\uff08a\uff09\u7b2c\u4e00\u9636\u6bb5\u6240\u5b66\u6a21\u578b\uff0c\u5373\u6d4b\u91cf\u6a21\u578b\uff1b\u5b50\u56fe\uff08b\uff09\u5b50\u56fe\uff08a\uff09\u7b2c\u4e8c\u9636\u6bb5\u6240\u5b66\u6a21\u578b\uff0c\u5373\u7ed3\u6784\u6a21\u578b\u3002<\/span><\/p>\n


<\/section>\n
\u8be5\u7814\u7a76\u8bc1\u660e\u4e86\u5982\u679c\u6bcf\u4e00\u4e2a\u9690\u53d8\u91cf\u81f3\u5c11\u6709\u4e09\u4e2a\u53ca\u5176\u4ee5\u4e0a\u7eaf\u7684\u6d4b\u91cf\u53d8\u91cf<\/span>\uff08\u8be5\u53d8\u91cf\u6709\u4e14\u53ea\u6709\u4e00\u4e2a\u9690\u53d8\u91cf\u4f5c\u4e3a\u5176\u7236\u8282\u70b9\uff09<\/span>\uff0c\u90a3\u4e48\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u7ed3\u6784\u53ef\u901a\u8fc7\u6d4b\u91cf\u53d8\u91cf\u53bb\u8bc6\u522b\uff0c\u8f93\u51fa\u7ed3\u679c\u8fbe\u5230Markov\u7b49\u4ef7\u7c7b\u3002\u4e4b\u540e\uff0c\u4e3a\u4e86\u9ad8\u6548\u5730\u5b66\u4e60\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u7ed3\u6784\uff0cKummerfeld\u7b49\u4eba[Kummerfeld et al., 2016]\u57fa\u4e8e\u66f4\u52a0\u5bbd\u6cdb\u7684\u4f4e\u79e9\u6761\u4ef6[Sullivant et al., 2010]\uff0c\u63d0\u51fa\u4e86FOFC<\/span>\uff08FindOneFactorClusters\uff09<\/span>\u7b97\u6cd5\u53bb\u4f30\u8ba1\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u7ed3\u6784\u3002<\/span><\/section>\n

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

(2) \u57fa\u4e8eGIN\u7684\u65b9\u6cd5<\/span><\/strong><\/h3>\n

<\/section>\n
\u4e0a\u4e00\u8282\u4e2d\u65b9\u6cd5\u5229\u7528\u7684\u662f\u53d8\u91cf\u7684\u534f\u65b9\u5dee\u77e9\u9635\u7684\u79e9\u7684\u7ea6\u675f<\/span>\uff08second-order statistics\uff09<\/span>\uff0c\u5ffd\u7565\u4e86\u53d8\u91cf\u4e2d\u9690\u542b\u7684\u975e\u9ad8\u65af\u6027<\/span> (High-order statistics)<\/span>\uff0c\u5bfc\u81f4\u90e8\u5206\u56e0\u679c\u4fe1\u606f\u88ab\u4e22\u5931\u3002\u4f8b\u5982\u56fe2\u4e2d\u7684\u56e0\u679c\u7ed3\u6784\uff0c\u5305\u542b4\u4e2a\u9690\u53d8\u91cf\u548c8\u4e2a\u89c2\u5bdf\u53d8\u91cf\uff0c\u4e0a\u8ff0\u7684\u65b9\u6cd5\u65e0\u6cd5\u4ece\u8fd98\u4e2a\u89c2\u5bdf\u53d8\u91cf\u4e2d\u53bb\u6062\u590d\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u7ed3\u6784\u3002\u5bf9\u4e8e\u53d8\u91cf\u7684\u975e\u9ad8\u65af\u800c\u8a00\uff0c\u9996\u5148\u6211\u4eec\u6ce8\u610f\u5230\u975e\u9ad8\u65af\u6027\u662f\u53ef\u4ee5\u6839\u636e\u89c2\u5bdf\u6570\u636e\u8f7b\u677e\u5730\u68c0\u9a8c\u6b64\u5047\u8bbe\u3002\u6b64\u5916\uff0c\u6b63\u5982Cram\u00e9r [1962]\u6240\u8ff0\u7684Cram\u00e9r\u5206\u89e3\u5b9a\u7406\uff0c\u4e0e\u9ad8\u65af\u53d8\u91cf\u4e0d\u540c\uff0c\u975e\u9ad8\u65af\u5206\u5e03\u7684\u53d8\u91cf\u9884\u8ba1\u5c06\u65e0\u5904\u4e0d\u5728\uff0cSpirtes\u548cZhang\u4e5f\u540c\u6837\u5728\u6587\u732e[Spirtes and Zhang 2016]\u4e2d\u6307\u51fa\u975e\u9ad8\u65af\u6570\u636e\u7684\u666e\u904d\u6027\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/section>\n
\u56fe2\uff1a\u6d89\u53ca4\u4e2a\u9690\u53d8\u91cf\u7684\u56e0\u679c\u7ed3\u6784\uff0c\u5176\u4e2dXi<\/sub>,i=1,\u2026,8\u4e3a\u89c2\u5bdf\u53d8\u91cf\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u4e3a\u6b64\uff0cXie\u7b49\u4eba[Xie et al., 2020 & Cai et al., 2019]\u6839\u636e\u89c2\u5bdf\u6570\u636e\u7684\u9ad8\u9636\u7edf\u8ba1\u91cf\uff0c\u8bbe\u8ba1\u4e86GIN<\/span> (Generalized Independent Noise condition) <\/span>\u6761\u4ef6\uff0c\u4ee5\u523b\u753b\u6f5c\u5728\u9690\u53d8\u91cf\u95f4\u7684\u5206\u79bb\u51c6\u5219\u3002\u76f4\u89c2\u4e0a\u6765\u8bf4\uff1a\u6cdb\u5316\u72ec\u7acb\u566a\u58f0<\/span>\uff08GIN\uff09<\/span> \u6761\u4ef6\u662f\u5173\u4e8e\u523b\u753b\u89c2\u5bdf\u53d8\u91cf\u7684\u7ec4\u5408\u4e4b\u95f4\u662f\u5426\u72ec\u7acb\u7684\u6027\u8d28\uff0c\u6240\u4ee5\u5b83\u53ef\u4ee5\u88ab\u8ba4\u4e3a\u662f\u7ecf\u5178\u7684\u72ec\u7acb\u566a\u58f0<\/span> (independent noise (IN)) <\/span>\u6761\u4ef6\u7684\u4e00\u79cd\u6269\u5c55\u3002\u6211\u4eec\u53d1\u73b0\uff0c\u901a\u8fc7\u68c0\u9a8c\u4e00\u4e9bGIN\u6761\u4ef6\u662f\u5426\u6210\u7acb\uff0c\u80fd\u591f\u5e2e\u52a9\u627e\u51fa\u56e0\u679c\u9690\u53d8\u91cf\u662f\u5426\u5b58\u5728\uff0c\u6240\u5728\u4f4d\u7f6e\uff0c\u4ee5\u53ca\u8fd9\u4e9b\u9690\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002\u57fa\u4e8e\u6b64\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u4e24\u9636\u6bb5\u7684GIN\u53d1\u73b0\u7b97\u6cd5\uff0c\u9996\u5148\u53bb\u5b9a\u4f4d\u9690\u53d8\u91cf\u53ca\u5176\u4e2a\u6570\uff0c\u5176\u6b21\u5b66\u4e60\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002<\/span><\/section>\n

<\/section>\n
\u4e3e\u4f8b\u6765\u8bf4\uff0c\u5982\u8003\u8651\u56fe2\u4e2d\uff0cGIN\u7b97\u6cd5\u5728\u7b2c\u4e00\u9636\u6bb5\u4e2d\u5b9a\u4f4d\u5230\u80cc\u540e\u5b58\u5728\u4e09\u7ec4\u9690\u53d8\u91cf{L1<\/sub>, L2<\/sub>}\uff0cL3<\/sub>\u548cL4<\/sub>\uff0c\u5176\u4e2d\u9690\u53d8\u91cf{L1<\/sub>, L2<\/sub>}\u5305\u542b\u7684\u5b50\u8282\u70b9\uff08X1<\/sub>, …, X4<\/sub>\uff09\u90fd\u76f8\u540c\u3002\u4e4b\u540e\u5728\u7b2c\u4e8c\u9636\u6bb5\u4e2d\u5b66\u4e60{L1<\/sub>, L2<\/sub>}\uff0cL3<\/sub>\u548cL4<\/sub>\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002<\/span><\/section>\n

<\/section>\n
\u4e0e\u4e0a\u4e00\u8282\u65b9\u6cd5\u76f8\u6bd4\uff0c\u8be5\u65b9\u6cd5\u660e\u663e\u7684\u4f18\u52bf\u662f\uff1a\u6bcf\u4e00\u4e2a\u9690\u53d8\u91cf\u6240\u9700\u7684\u7eaf\u7684\u6d4b\u91cf\u53d8\u91cf\u66f4\u5c11\uff0c\u4e14\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u65b9\u5411\u90fd\u662f\u53ef\u8bc6\u522b\u7684\u3002\u6700\u8fd1Chen\u7b49\u4eba[Chen et al., 2022]\uff0c\u8003\u8651\u566a\u58f0\u662f\u4efb\u610f\u5206\u5e03\u60c5\u51b5\u4e0b\uff0c\u5206\u6790\u5e76\u7ed9\u51fa\u4e86\u7b2c\u4e8c\u6b65\u4e2d\u53d8\u91cf\u95f4\u56e0\u679c\u65b9\u5411\u8bc6\u522b\u7684\u5145\u5206\u5fc5\u8981\u6761\u4ef6\u3002<\/span><\/section>\n

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

(3) \u5c42\u7ea7\u7ed3\u6784\u5b66\u4e60\u7684\u65b9\u6cd5\uff08latent hierarchical structure\uff09<\/span><\/strong><\/h3>\n

<\/section>\n
\u672c\u5c0f\u8282\u7ee7\u7eed\u8003\u8651\u53d1\u73b0\u9690\u85cf\u7684\u56e0\u679c\u53d8\u91cf\u53ca\u5176\u5b83\u4eec\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u95ee\u9898\u3002\u76f8\u8f83\u4e8e\u4e0a\u8ff0\u4e24\u4e2a\u5c0f\u8282\u7684\u6a21\u578b\uff0c\u8fd9\u91cc\u8003\u8651\u66f4\u52a0\u5bbd\u6cdb\u7684\u4e00\u79cd\u6a21\u578b\u8bbe\u5b9a\uff0c\u5141\u8bb8\u4e00\u4e9b\u9690\u85cf\u7684\u53d8\u91cf\u6ca1\u6709\u76f4\u63a5\u89c2\u5bdf\u7684\u53d8\u91cf\u4f5c\u4e3a\u5176\u5b50\u8282\u70b9\uff0c\u5982\u56fe3\u4e2d\u7684\u9690\u53d8\u91cfL1<\/sub>\u3002\u4e3a\u89e3\u51b3\u8be5\u95ee\u9898\uff0c\u6700\u8fd1Xie\u7b49\u4eba[Xie et al., 2022]\u63d0\u51fa\u4e86\u4e00\u79cd\u6700\u5c0f\u5c42\u7ea7\u7ed3\u6784\u7684\u5145\u5206\u6761\u4ef6\uff0c\u5e76\u8bc1\u660e\u4e86\u5728\u8be5\u6761\u4ef6\u4e0b\uff0c\u7ebf\u6027\u975e\u9ad8\u65af\u7684\u9690\u53d8\u91cf\u5c42\u7ea7\u7ed3\u6784\u662f\u53ef\u4ee5\u901a\u8fc7\u90e8\u5206\u89c2\u5bdf\u53d8\u91cf\u5b8c\u5168\u8bc6\u522b\u7684\uff0c\u5305\u62ec\u9690\u53d8\u91cf\u7684\u4e2a\u6570\u53ca\u5176\u56e0\u679c\u7ed3\u6784\u5173\u7cfb\u3002<\/span><\/section>\n

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

\u56fe3\uff1a\u4e00\u4e2a\u6d89\u53ca9\u4e2a\u9690\u53d8\u91cf(Li<\/sub>, i=1,\u2026,9)\u548c15\u4e2a\u89c2\u5bdf\u53d8\u91cf(Xi<\/sub>, i=1,\u2026,15)\u7684\u5c42\u7ea7\u56e0\u679c\u7ed3\u6784\u3002<\/span><\/p>\n


<\/section>\n
\u5177\u4f53\u6765\u8bb2\uff0c\u8be5\u7814\u7a76\u9996\u5148\u5b9a\u4e49\u4e86\u4e00\u79cd\u5145\u5206\u7684\u6700\u5c0f\u9690\u53d8\u91cf\u5c42\u7ea7\u7ed3\u6784\u6761\u4ef6<\/span>\uff08minimal latent hierarchical structure condition\uff09<\/span>\uff1a(1) \u6bcf\u4e00\u4e2a\u9690\u53d8\u91cf\u81f3\u5c11\u6709\u4e09\u4e2a\u90bb\u5c45\u8282\u70b9\uff1b(2) \u6bcf\u4e00\u4e2a\u9690\u53d8\u91cf\u81f3\u5c11\u6709\u4e24\u4e2a\u7eaf\u7684\u5b50\u8282\u70b9<\/span>\uff08\u8be5\u8282\u70b9\u53ef\u4ee5\u662f\u9690\u7684\u4e5f\u53ef\u4ee5\u662f\u89c2\u5bdf\u7684\uff09<\/span>\u3002\u7814\u7a76\u8bc1\u660e\uff1a\u5728\u7ebf\u6027\u975e\u9ad8\u65af\u7cfb\u7edf\u4e0b\uff0c\u5982\u679c\u6700\u5c0f\u9690\u53d8\u91cf\u5c42\u7ea7\u7ed3\u6784\u6761\u4ef6\u6210\u7acb\uff0c\u90a3\u4e48\u7cfb\u7edf\u80cc\u540e\u6240\u6709\u7684\u9690\u53d8\u91cf\u53ca\u5176\u56e0\u679c\u5173\u7cfb\u53ef\u4ee5\u4ec5\u6709\u7684\u89c2\u5bdf\u53d8\u91cf\u53bb\u6062\u590d\u3002\u7814\u7a76\u63d0\u51fa\u7684LaHME\u7b97\u6cd5\u5305\u542b\u4ee5\u4e0b\u4e24\u4e2a\u6b65\u9aa4\uff1a\u6b65\u9aa4\u4e00\uff1a\u8fed\u4ee3\u5730\u5b9a\u4f4d\u6240\u6709\u6f5c\u5728\u9690\u53d8\u91cf\u7684\u8fc7\u7a0b\uff0c\u5305\u62ec(i) \u4ece\u5f53\u524d\u6d3b\u52a8\u53d8\u91cf\u96c6\u4e2d\u8bc6\u522b\u56e0\u679c\u7c7b<\/span> (\u5171\u4eab\u5171\u540c\u7236\u8282\u70b9\u7684\u96c6\u5408)<\/span>\uff1b(ii) \u786e\u5b9a\u9700\u8981\u4e3a\u8fd9\u4e9b\u56e0\u679c\u7c7b\u5f15\u5165\u7684\u65b0\u6f5c\u5728\u53d8\u91cf\u7684\u6570\u91cf\uff1b(iii) \u66f4\u65b0\u6d3b\u52a8\u53d8\u91cf\u96c6\u4e09\u4e2a\u5b50\u8fc7\u7a0b\u3002\u6b65\u9aa4\u4e8c\uff1a\u63a8\u65ad\u5df2\u8bc6\u522b\u6f5c\u5728\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u7ed3\u6784\uff0c\u5305\u62ec\u8bc6\u522b\u6f5c\u5728\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u987a\u5e8f\u548c\u53bb\u9664\u591a\u4f59\u7684\u8fb9\u4e24\u4e2a\u5b50\u8fc7\u7a0b\u3002\u5982\u8003\u8651\u56fe3\uff0c\u5728LaHME\u7b97\u6cd5\u6b65\u9aa4\u4e00\u4e2d\uff0c\u7b2c\u4e00\u6b21\u8fed\u4ee3\u6211\u4eec\u4f1a\u53d1\u73b0\u7ed3\u6784\u4e2d\u6700\u4e0b\u9762\u4e00\u5c42\u76846\u4e2a\u9690\u53d8\u91cf\uff0c\u5373L4<\/sub>, …, L9<\/sub>\u3002\u4e4b\u540e\u6211\u4eec\u8fed\u4ee3\u5730\u627e\u51fa\u5269\u4f59\u7684\u9690\u53d8\u91cf\u3002LaHME\u7b97\u6cd5\u7684\u6b65\u9aa4\u4e8c\u4e2d\uff0c\u6211\u4eec\u9700\u8981\u8fdb\u4e00\u6b65\u786e\u5b9a\u9690\u53d8\u91cf\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u5982L2<\/sub>\u4e0eL5<\/sub>\uff0cL3<\/sub>\u4e0eL8<\/sub>\u7684\u5173\u7cfb\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff1a\u5982\u679c\u80cc\u540e\u7684\u56e0\u679c\u7ed3\u6784\u662f\u4e00\u79cd\u7ecf\u5178\u7684Tree-based\u7ed3\u6784\uff0c\u90a3\u4e48\u5229\u7528LaHME\u7b97\u6cd5\u4e2d\u7684\u6b65\u9aa4\u4e00\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u5b8c\u5168\u6062\u590d\u80cc\u540e\u7684\u56e0\u679c\u56fe\uff0c\u4e0d\u9700\u8981\u8fdb\u884c\u6b65\u9aa4\u4e8c\u7684\u64cd\u4f5c\u3002<\/span><\/section>\n

<\/section>\n
\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u672c\u7814\u7a76\u4e0d\u9700\u8981\u63d0\u524d\u5df2\u77e5\u7cfb\u7edf\u80cc\u540e\u9690\u53d8\u91cf\u7684\u4e2a\u6570\u53ca\u5176\u5c42\u7ea7\u7ed3\u6784\u7684\u5c42\u6570\u3002\u5982\u679c\u6211\u4eec\u63d0\u524d\u5df2\u77e5\u80cc\u540e\u9690\u53d8\u91cf\u7684\u4e2a\u6570\uff0c\u90a3\u4e48\u6839\u636eAdams\u7b49\u4eba[Adams et al., 2021]\u7ed9\u51fa\u7684\u5145\u5206\u7684\u5fc5\u8981\u53ef\u8bc6\u522b\u6027\u6761\u4ef6\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u90e8\u5206\u89c2\u5bdf\u53d8\u91cf\uff0c\u5229\u7528\u5b8c\u5907\u7684ICA\u7b97\u6cd5\u6062\u590d\u5176\u80cc\u540e\u7684\u9690\u53d8\u91cf\u5c42\u7ea7\u7ed3\u6784\u3002\u6700\u8fd1Huang\u7b49\u4eba[Huang et al., 2022]\u5229\u7528\u6570\u636e\u534f\u65b9\u5dee\u77e9\u9635\u7684rank\u7ea6\u675f<\/span>\uff08\u4e8c\u9636\u7edf\u8ba1\u91cf\uff09<\/span>\uff0c\u8bc1\u660e\u4e86\u5728\u4e00\u4e9b\u56fe\u7ed3\u6784\u7684\u9650\u5236\u4e0b\uff0c\u6240\u63d0\u51fa\u7684\u7b97\u6cd5\u53ef\u4ee5\u6e10\u8fd1\u5730\u627e\u5230\u9690\u5c42\u7ea7\u7ed3\u6784\u6240\u5bf9\u5e94\u7684\u9a6c\u5c14\u53ef\u592b\u7b49\u4ef7\u7c7b\u3002\u76f8\u8f83\u4e8e\u4e0a\u8ff0GIN-based\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u5141\u8bb8\u7cfb\u7edf\u542b\u6709multi-factor\u7684\u60c5\u5f62\uff0c\u5373\u591a\u4e2a\u8282\u70b9\u5171\u4eab\u4e24\u4e2a\u6216\u8005\u4ee5\u4e0a\u7684\u9690\u8282\u70b9\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u6211\u4eec\u77e5\u9053\u56e0\u679c\u5173\u7cfb\u5bf9\u6270\u52a8\u5177\u6709\u9c81\u68d2\u6027\uff0c\u5177\u6709\u7f16\u7801\u7cfb\u7edf\u4e2d\u7684\u4e0d\u53d8\u6027\uff0c\u4ece\u800c\u80fd\u591f\u6709\u6548\u5730\u63a8\u7406\u5176\u884c\u4e3a\u5728\u73af\u5883\u4e2d\u7684\u5f71\u54cd\u3002\u8be5\u7814\u7a76\u5bf9\u4e00\u4e9b\u673a\u5668\u5b66\u4e60\u7684\u4efb\u52a1\uff0c\u5982\u56fe\u7247\u8bc6\u522b\u95ee\u9898\uff0c\u6709\u91cd\u8981\u7684\u5f71\u54cd\u610f\u4e49\u3002\u8fd9\u662f\u56e0\u4e3a\u5f80\u5f80\u4ea7\u751f\u56e0\u679c\u5173\u7cfb\u7684\u53d8\u91cf\u662f\u4e00\u4e9b\u7684\u9ad8\u7ea7\u7684\u9690\u7684\u56e0\u679c\u7279\u5f81\uff0c\u800c\u975e\u76f4\u63a5\u89c2\u5bdf\u7684\u53d8\u91cf\uff0c\u5982\u50cf\u7d20\u3002\u56e0\u6b64\u95ee\u9898\u5c31\u53d8\u6210\u4e86\u627e\u5230\u80cc\u540e\u771f\u6b63\u53d1\u751f\u56e0\u679c\u5173\u7cfb\u7684\u56e0\u679c\u7279\u5f81\uff0c\u4ece\u800c\u63a8\u65ad\u8fd9\u4e9b\u7279\u5f81\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u5b66\u4e60\u9690\u53d8\u91cf\u5c42\u7ea7\u7ed3\u6784\u7684\u95ee\u9898\u3002<\/span><\/section>\n

<\/section>\n

(4) \u5c0f\u7ed3<\/span><\/strong><\/span><\/h3>\n

<\/section>\n
\u8fd9\u4e00\u7ae0\u8282\u7ed9\u51fa\u4e86\u5982\u4f55\u4ece\u5c11\u91cf\u7684\u89c2\u5bdf\u6570\u636e\u4e2d\u5b66\u4e60\u7ebf\u6027\u7cfb\u7edf\u4e0b\u7684\u9690\u53d8\u91cf\u56e0\u679c\u7ed3\u6784\u95ee\u9898\uff0c\u7ed9\u51fa\u4e86\u5176\u7ed3\u6784\u53ef\u8bc6\u522b\u6027\u7684\u6761\u4ef6\u548c\u4f30\u8ba1\u7b97\u6cd5\u3002\u7136\u800c\u76ee\u524d\u7684\u7814\u7a76\u4ecd\u7136\u5b58\u5728\u4e00\u4e9b\u4e0d\u8db3\uff0c\u5982\u9700\u8981\u5047\u5b9a\u6240\u6709\u7684\u89c2\u5bdf\u53d8\u91cf\u4e0d\u80fd\u591f\u5f71\u54cd\u9690\u53d8\u91cf\uff0c\u5373\u6d4b\u91cf\u5047\u8bbe\u3002\u7b2c\u4e8c\u4e2a\u4e0d\u8db3\u662f\u7ed3\u8bba\u662f\u4ec5\u9650\u4e8e\u7ebf\u6027\u7cfb\u7edf\u4e0b\u7684\u9690\u53d8\u91cf\u7ed3\u6784\u5b66\u4e60\uff0c\u5982\u4f55\u4f30\u8ba1\u975e\u7ebf\u6027\u56e0\u679c\u6a21\u578b\u4e0b\u7684\u6f5c\u5728\u9690\u7ed3\u6784\u5e76\u672a\u7ed9\u51fa\u3002\u4e00\u4e2a\u53ef\u884c\u7684\u65b9\u6cd5\u662f\u501f\u9274\u7ecf\u5178\u7684\u975e\u7ebf\u6027\u6a21\u578b\u6846\u67b6\uff0c\u5982\u540e\u975e\u7ebf\u6027\u6a21\u578b [Zhang & Hyv\u00e4rinen, 2009]\u53bb\u63a2\u7d22\u89e3\u51b3\u3002<\/span><\/section>\n

<\/section>\n

2.2 \u65f6\u95f4\u5e8f\u5217<\/span><\/strong><\/span><\/h2>\n


<\/h3>\n

(1) \u4f20\u7edf\u56e0\u679c\u53d1\u73b0: Granger causality<\/span><\/strong><\/h3>\n

<\/section>\n
\u683c\u5170\u6770\u56e0\u679c\u5173\u7cfb\u68c0\u9a8c<\/span>\uff08Granger causality test\uff09<\/span>\u662f\u4e00\u79cd\u5728\u65e0\u5e72\u6270\u56e0\u7d20\u7684\u60c5\u51b5\u4e0b\uff0c\u7528\u6765\u5bfb\u627e\u65f6\u5e8f\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u7684\u65b9\u6cd5\u3002Granger causality test\u75312003\u5e74\u8bfa\u8d1d\u5c14\u7ecf\u6d4e\u5b66\u5956\u5f97\u4e3bClive W. J. Granger\u6240\u5f00\u521b\uff0c\u5e38\u53ef\u7528\u4e8e\u5206\u6790\u7ecf\u6d4e\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002\u683c\u5170\u6770\u56e0\u679c\u5173\u7cfb\u68c0\u9a8c\u7684\u57fa\u672c\u89c2\u5ff5\u5728\u4e8e\uff1a\uff081\uff09 \u672a\u6765\u7684\u4e8b\u4ef6\u4e0d\u4f1a\u5bf9\u76ee\u524d\u4e0e\u8fc7\u53bb\u4ea7\u751f\u56e0\u679c\u5f71\u54cd\uff0c\u800c\u8fc7\u53bb\u7684\u4e8b\u4ef6\u624d\u53ef\u80fd\u5bf9\u73b0\u5728\u53ca\u672a\u6765\u4ea7\u751f\u5f71\u54cd\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u5982\u679c\u6211\u4eec\u8bd5\u56fe\u63a2\u8ba8\u53d8\u91cfX\u662f\u5426\u5bf9\u53d8\u91cfY\u6709\u56e0\u679c\u5f71\u54cd\uff0c\u90a3\u4e48\u53ea\u9700\u8981\u4f30\u8ba1x\u7684\u843d\u540e\u671f\u662f\u5426\u4f1a\u5f71\u54cdy\u7684\u73b0\u5728\u503c\u548c\uff082\uff09cause variable\u9700\u8981\u5305\u542beffect variable\u91cc\u72ec\u7279\u975e\u5197\u4f59\u7684\u4fe1\u606f\u3002\u5047\u5982\u5728\u63a7\u5236\u4e86Y\u53d8\u91cf\u7684\u8fc7\u53bb\u503c\u4ee5\u540e\uff0cx \u53d8\u91cf\u7684\u8fc7\u53bb\u503c\u4ecd\u80fd\u5bf9Y \u53d8\u91cf\u6709\u663e\u8457\u7684\u89e3\u91ca\u80fd\u529b\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u79f0X\u80fd\u201cGranger \u5f71\u54cd\u201d<\/span>\uff08Granger-cause) <\/span>y\u3002<\/span><\/section>\n

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

<\/section>\n
\u5728\u7b97\u6cd5\u5b9e\u73b0\u4e0a\uff0c\u6211\u4eec\u53ef\u901a\u8fc7\u975e\u53c2\u6570\u4f30\u8ba1<\/span>\uff08Nonparametric\uff09<\/span>\u7684\u65b9\u5f0f\uff0c\u6bd4\u5982\u5728PC algorithm\u4e0a\u52a0\u4e0a\u65f6\u5e8f\u9650\u5236\uff0c\u5373\u53ea\u5141\u8bb8\u56e0\u679c\u65b9\u5411\u4ece\u8fc7\u53bb\u6307\u5411\u672a\u6765\u5e76\u5728\u505a\u72ec\u7acb\u68c0\u9a8c\u65f6\u53bb\u9664\u672a\u6765\u4fe1\u606f\u7684\u65b9\u5f0f\u6765\u6dfb\u52a0Granger temporal constraints\u3002\u6211\u4eec\u901a\u5e38\u7528\u7ebf\u6027\u81ea\u56de\u5f52\u6a21\u578bVector autoregessive model\u8fdb\u884c\u7b80\u5316\u3002<\/span><\/section>\n

<\/section>\n
\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u4f7f\u7528Granger causality test\u5bfb\u627e\u56e0\u679c\u5173\u7cfb\u65f6\uff0c\u6211\u4eec\u5fc5\u987b\u4f7f\u7528\u7ed3\u6784\u5316\u6570\u636e\uff0c\u5373\u6570\u636e\u4e2d\u76f4\u63a5\u542b\u6709\u6240\u6709\u56e0\u679c\u53d8\u91cf<\/span>\uff08causal variable\uff09<\/span>\u3002\u5f53\u89c2\u6d4b\u6570\u636e\u4e2d\u7684\u53d8\u91cf\u4e0d\u5b58\u5728\u76f4\u63a5\u56e0\u679c\u5173\u7cfb\uff0c\u800c\u662f\u7531\u56e0\u679c\u9690\u53d8\u91cf\u6216\u8005\u5e72\u6270\u56e0\u7d20<\/span>\uff08counfounder\uff09<\/span>\u4ea7\u6210\u7684\uff0c\u5982\u89c6\u9891\u6570\u636e\uff0cGranger causality test\u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528\u3002<\/span><\/section>\n

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

(2) \u4ece\u9759\u6001\u65f6\u95f4\u5e8f\u5217\u4e2d\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb<\/span><\/strong><\/h3>\n

<\/section>\n
\u5728\u4e0a\u4e00\u7ae0\u8282\uff0c\u6211\u4eec\u8ba8\u8bba\u4e86\u4f20\u7edf\u56e0\u679c\u53d1\u73b0\u7684\u65b9\u6cd5\u7684\u9002\u7528\u573a\u666f\u3002\u7136\u800c\uff0c\u5982\u679c\u6211\u4eec\u60f3\u4ece\u5927\u591a\u6570\u5b9e\u9645\u65f6\u5e8f\u4fe1\u53f7\u6570\u636e\uff0c\u5982\u89c6\u9891\u6570\u636e\uff0c\u4e2d\u5bfb\u627e\u56e0\u679c\u5173\u7cfb\uff0cGranger causality\u662f\u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528\u7684\u3002\u5728\u8fd9\u79cd\u60c5\u5f62\u4e0b\uff0c\u6211\u4eec\u89c2\u6d4b\u5230\u7684\u65f6\u5e8f\u53d8\u91cf\uff0c\u5373\u89c6\u9891\u4e2d\u6bcf\u5e27\u7684\u50cf\u7d20\uff0c\u5b83\u4eec\u4e4b\u95f4\u5e76\u4e0d\u5b58\u5728\u76f4\u63a5\u7684\u56e0\u679c\u5173\u7cfb\u3002\u8fd9\u4e9b\u65f6\u5e8f\u53d8\u91cf<\/span>\uff08\u50cf\u7d20\uff09<\/span>\u5f80\u5f80\u662f\u7531\u5177\u6709\u65f6\u5e8f\u56e0\u679c\u5173\u7cfb\u7684\u56e0\u679c\u9690\u53d8\u91cf\u6216\u5e72\u6270\u56e0\u7d20\u751f\u6210\u7684\u3002\u5728\u8fd9\u79cd\u60c5\u578b\u4e0b\uff0c\u6211\u4eec\u9700\u8981\u505a\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u4ece\u65f6\u95f4\u5e8f\u5217\u4e2d\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\uff0c\u5e76\u4e14\u8bc6\u522b\u5b83\u4eec\u4e4b\u95f4\u7684\u65f6\u5e8f\u56e0\u679c\u5173\u7cfb\u3002\u5728\u4e00\u822c\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u8868\u5f81\u5b66\u4e60\u5b66\u5f97\u4f4e\u7ef4\u7684\u6570\u636e\u8868\u5f81\uff0c\u4f46\u6211\u4eec\u65e0\u6cd5\u4fdd\u8bc1\u80fd\u6062\u590d\u771f\u6b63\u7684\u56e0\u679c\u9690\u53d8\u91cf\u3002\u5728\u672c\u7ae0\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\u5728[Yao et al., 2021 & 2022]\u4e2d\u63d0\u51fa\u7684\u4e24\u79cd\u4ece\u9759\u6001\u65f6\u95f4\u5e8f\u5217\u4e2d\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u56e0\u679c\u5173\u7cfb\u7684\u65b9\u6cd5\u3002<\/span><\/section>\n


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

Linear Latent Causal Processes with Generalized Laplacian Noise<\/span><\/em><\/strong><\/h4>\n

<\/section>\n
\u5728\u4f20\u7edf\u56e0\u679c\u53d1\u73b0\u4e2d\uff0cLinear Non-Gaussian\u6a21\u578b\u901a\u5e38\u662f\u6211\u4eec\u671f\u671b\u7684\u6a21\u578b\u5047\u8bbe\u3002\u5728\u8fd9\u79cd\u5047\u8bbe\u4e0b\uff0c\u6211\u4eec\u901a\u5e38\u53ef\u4ee5\u5b9e\u73b0\u53ef\u8bc6\u522b\uff0c\u6709\u4fdd\u969c\u7684\u56e0\u679c\u53d1\u73b0\u3002\u6211\u4eec\u9996\u5148\u5c06\u8fd9\u79cdInductive Bias\u52a0\u5728Latent\u7a7a\u95f4 [Yao et al., 2021]\uff0c\u5047\u8bbe\u5b9e\u9645\u7684\u9690\u53d8\u91cf\u4e0e\u5176\u843d\u540e\u671f\u4e4b\u95f4\u6ee1\u8db3\u7ebf\u6027\u5047\u8bbe<\/span>\uff08LNM\uff09<\/span>\u3002\u89c2\u6d4b\u5230\u7684\u6570\u636eXt<\/sub>\u5219\u662f\u9690\u53d8\u91cfZt<\/sub>\u7684\u975e\u7ebf\u6027<\/span>\uff08\u4f46\u53ef\u9006\uff09<\/span>\u7684\u6620\u5c04\u3002\u53e6\u5916\uff0c\u53ea\u8981\u6211\u4eec\u5047\u8bbe\u56e0\u679c\u8fc7\u7a0b\u4e2d\u7684\u566a\u58f0<\/span>\uff08Process Noise\uff09<\/span>\u4e4b\u95f4\u4e92\u76f8\u72ec\u7acb\uff0c\u4e14\u6ee1\u8db3Genenralized Laplacian\u5206\u5e03<\/span>\uff08\u4e00\u79cdNon-Gaussian\u6a21\u578b\u5047\u8bbe\uff09<\/span>\u3002\u901a\u8fc7\u7406\u8bba\u63a8\u5bfc\uff0c\u6211\u4eec\u53d1\u73b0\u53ea\u8981\u5c06\u8fd9\u4e9b\u5047\u8bbe\u8f6c\u6362\u4e3aVariational Autoencoder <\/span>\uff08VAE\uff09<\/span>\u7684\u53c2\u6570\u5316\u7684\u65b9\u5f0f\uff0c\u6211\u4eec\u5c31\u53ef\u4ee5\u6709\u4fdd\u8bc1\u5730\u6062\u590d\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002<\/span><\/section>\n

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

Linear Latent Causal Processes with Generalized Laplacian Noise\u6a21\u578b\u5047\u8bbe<\/span><\/p>\n


<\/section>\n
\u6211\u4eec\u9996\u5148\u5728KiTTiMask\u6570\u636e\u4e0a\u9a8c\u8bc1\u6211\u4eec\u7684\u7406\u8bba\u7684\u6a21\u578b\u3002KiTTiMask\u662f\u4e00\u4e2a\u8bb0\u5f55\u884c\u4eba\u6b65\u884c\u7684\u89c6\u9891\u6570\u636e\u96c6\uff0c\u5176\u5305\u542b\u7684\u4e09\u4e2a\u56e0\u679c\u9690\u53d8\u91cf\uff0c\u5373\u884c\u4eba\u5728\u89c6\u9891\u4e2d\u7684\u6a2a\u5411\u4f4d\u7f6e\uff0c\u7eb5\u5411\u4f4d\u7f6e\u548c\u884c\u4eba\u5728\u89c6\u9891\u4e2d\u7684\u5927\u5c0f\uff0c\u5b83\u4eec\u4e4b\u95f4\u6ee1\u8db3\u72ec\u7acb\u7ebf\u6027\u65f6\u5e8f\u5173\u7cfb\u5e76\u4e14\u566a\u58f0\u53ef\u8fd1\u4f3c\u6210Laplacian Noise\u3002\u4e0b\u56fe\u4e2d\uff0c\u5b9e\u9a8c\u7ed3\u679c\u9a8c\u8bc1\u4e86\u6211\u4eec\u7684\u6a21\u578b\u80fd\u5728\u72ec\u7acb\u7ebf\u6027\u65f6\u5e8f\u5173\u7cfb\u7684\u65f6\u5e8f\u6570\u636e\u4e2d\u6062\u590d\u51fa\u56e0\u679c\u9690\u53d8\u91cf\u3002<\/span><\/section>\n

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

KiTTiMask\u5b9e\u9a8c\u7ed3\u679c\uff1a\uff08a\uff09\u6062\u590d\u7684\u9690\u53d8\u91cf\u548c\u771f\u5b9e\u56e0\u679c\u9690\u53d8\u91cf\u4e4b\u95f4\u7684\u76f8\u5173\u7cfb\u6570\uff1b\uff08b\uff09\u6062\u590d\u7684\u56e0\u53d8\u91cf\u548c\u771f\u5b9e\u56e0\u679c\u9690\u53d8\u91cf\u4e4b\u95f4\u7684\u6563\u70b9\u56fe\uff1b\uff08c\uff09\u6062\u590d\u51fa\u6765\u7684\u56e0\u679c\u5173\u7cfb\u77e9\u9635\uff1b\uff08d\uff09Latent traveral\u5c55\u793a\u56e0\u679c\u9690\u53d8\u91cf\u5982\u4f55\u5f71\u54cd\u56fe\u50cf\u5185\u5bb9\u3002<\/span><\/p>\n


<\/section>\n
\u4e0b\u4e00\u6b65\uff0c\u6211\u4eec\u5728\u4e8c\u7ef4\u8d28\u91cf\u5f39\u7c27\u7cfb\u7edf\u89c6\u9891\u6570\u636e\u4e0a\u9a8c\u8bc1\u6211\u4eec\u7684\u7406\u8bba\u548c\u6a21\u578b\u3002\u8d28\u91cf\u5f39\u7c27\u7cfb\u7edf\u6ee1\u8db3\u7ebf\u6027\u5047\u8bbe\u3002\u6211\u4eec\u4eba\u4e3a\u7684\u5728\u6570\u636e\u751f\u6210\u8fc7\u7a0b\u4e2d\u5728\u6a2a\u5411\u548c\u7eb5\u5411\u4e0a\u52a0\u4e0a\u72ec\u7acb\u7684Laplacian Noise\uff0c\u4f7f\u751f\u6210\u7684\u6570\u636e\u6ee1\u8db3\u6211\u4eec\u7684\u6761\u4ef6\u3002\u4e0b\u56fe\u4e2d\uff0c\u5b9e\u9a8c\u7ed3\u679c\u9a8c\u8bc1\u4e86\u6211\u4eec\u7684\u6a21\u578b\u80fd\u5728\u7ebf\u6027\u65f6\u5e8f\u5173\u7cfb\u7684\u65f6\u5e8f\u6570\u636e\u4e2d\u6062\u590d\u51fa\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u3002<\/span><\/section>\n

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

\u8d28\u91cf\u5f39\u7c27\u7cfb\u7edf\u89c6\u9891\u6570\u636e\u5b9e\u9a8c\u7ed3\u679c\uff1a\uff08a\uff09\u6062\u590d\u7684\u9690\u53d8\u91cf\u548c\u771f\u5b9e\u56e0\u679c\u9690\u53d8\u91cf\u4e4b\u95f4\u7684\u76f8\u5173\u7cfb\u6570\uff1b\uff08b\uff09\u6062\u590d\u51fa\u6765\u7684\u56e0\u679c\u5173\u7cfb\u77e9\u9635\u3002<\/span><\/p>\n


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

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


<\/h4>\n

Nonparametric Latent Causal Processes<\/span><\/em><\/strong><\/h4>\n

<\/section>\n
\u5728\u4e0a\u5c0f\u8282\u4e2d\uff0c\u6211\u4eec\u63a2\u8ba8\u4e86\u5728\u65f6\u5e8f\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5176\u843d\u540e\u671f\u4e4b\u95f4\u6ee1\u8db3\u7ebf\u6027\u5173\u7cfb\u65f6\uff0c\u5982\u4f55\u901a\u8fc7\u6a21\u578b\u5047\u8bbe\u548c\u6dfb\u52a0\u5f52\u7eb3\u504f\u7f6e\u7684\u65b9\u5f0f\u6062\u590d\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u7136\u800c\uff0c\u7ebf\u6027\u5047\u8bbe\u662f\u4e00\u79cd\u5f88\u5f3a\u7684\u51fd\u6570\u5f62\u5f0f\u5047\u8bbe\uff0c\u8bb8\u591a\u65f6\u5e8f\u6570\u636e\u5e76\u4e0d\u6ee1\u8db3\u9690\u53d8\u91cf\u4e4b\u95f4\u7684\u7ebf\u6027\u5047\u8bbe\u3002\u5728\u8fd9\u4e2a\u5c0f\u8282\uff0c\u6211\u4eec\u7ee7\u7eed\u8ba8\u8bba\u5728\u56e0\u679c\u9690\u53d8\u91cf\u4e4b\u95f4\u6ee1\u8db3\u975e\u7ebf\u6027\uff0c\u751a\u81f3\u975e\u53c2\u6570\u5316<\/span>\uff08Nonparametric\uff09<\/span>\u7684\u5f62\u5f0f\u4e0b\uff0c\u5982\u4f55\u5b9e\u73b0\u56e0\u679c\u8868\u5f81\u5b66\u4e60 \u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/section>\n
Nonparametric Latent Causal Processes\u6a21\u578b\u5047\u8bbe<\/span><\/section>\n

<\/section>\n
\u4e0e\u4e0a\u5c0f\u8282\u76f8\u540c\uff0c\u6211\u4eec\u4ecd\u5047\u8bbe\u6570\u636eXt<\/sub>\u5219\u662f\u9690\u53d8\u91cfZt<\/sub>\u7684\u975e\u7ebf\u6027<\/span>\uff08\u4f46\u53ef\u9006\uff09<\/span>\u7684\u6620\u5c04\u3002\u4e0d\u540c\u7684\u662f\uff0c\u6211\u4eec\u4e0d\u518d\u5bf9\u65f6\u5e8f\u56e0\u679c\u9690\u53d8\u91cf\u4e4b\u95f4\u7684\u5173\u7cfb\u6216\u566a\u58f0\u5206\u5e03\u5f62\u5f0f\u505a\u4efb\u4f55\u53c2\u6570\u5316\u5047\u8bbe\u3002\u7136\u800c\uff0c\u5929\u4e0b\u6ca1\u6709\u514d\u8d39\u7684\u5348\u9910\u3002\u975e\u53c2\u6570\u5316<\/span>\uff08Nonparametric\uff09<\/span>\u7684\u5047\u8bbe\u5fc5\u7136\u9700\u8981\u522b\u7684\u4ee3\u4ef7\u548c\u6761\u4ef6\u6765\u8865\u507f\u3002\u901a\u8fc7\u7406\u8bba\u63a8\u5bfc [Yao et al., 2022]\uff0c\u6211\u4eec\u53d1\u73b0\u5728\u5f15\u5165\u4e86\u56e0\u679c\u5173\u7cfb\u6307\u6570\u5bfc\u6570\u4e4b\u95f4\u7684\u7ebf\u6027\u72ec\u7acb\u6761\u4ef6<\/span>\uff08Linear Independence Condition\uff09<\/span>\u540e\uff0c\u9664\u4e86Linear Gaussian Model\u4e4b\u5916\u7684Latent Process\uff0c\u90fd\u6709\u6781\u5927\u53ef\u80fd\u6ee1\u8db3\u8be5\u7ebf\u6027\u72ec\u7acb\u6761\u4ef6\u3002\u8be5\u6761\u4ef6\u53ef\u5e7f\u6cdb\u4f7f\u7528\u4e8e\u65f6\u5e8f\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u5728\u975e\u53c2\u60c5\u51b5\u4e0b\u5bfb\u627e\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb\u3002<\/span><\/section>\n

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

(3) \u6570\u636e\u5206\u5e03\u53d8\u5316\u4e3a\u65f6\u95f4\u5e8f\u5217\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u5e26\u6765\u7684\u597d\u5904<\/span><\/strong><\/h3>\n

<\/span><\/section>\n
\u5728\u4e0a\u4e00\u7ae0\u8282\u4e2d\uff0c\u6211\u4eec\u8ba8\u8bba\u4e86\u5728\u9759\u6001\u65f6\u5e8f\u73af\u5883\u4e0b\u7684\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u3002\u4e0b\u4e00\u6b65\uff0c\u6211\u4eec\u5c06\u7ee7\u7eed\u63a2\u8ba8\u5728\u6570\u636e\u5206\u5e03\u53d1\u751f\u53d8\u5316\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u8be5\u5982\u4f55\u6062\u590d\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6211\u4eec\u7684\u5206\u6790\u662f\u5728\u80fd\u63a2\u6d4b\u5230\u5206\u5e03\u53d8\u5316\u7684\u5047\u8bbe\u4e0b\u8fdb\u884c\u7684\u3002\u5373\u5728\u6062\u590d\u9690\u53d8\u91cf\u8fc7\u7a0b\u524d\uff0c\u6211\u4eec\u5df2\u77e5\u4f55\u65f6\u6570\u636e\u53d1\u751f\u5206\u5e03\u53d8\u5316\u548c\u54ea\u79cd\u7c7b\u522b<\/span>\uff08class\uff09<\/span>\u7684\u53d8\u5316\uff0c\u6bd4\u5982\u57df\u7d22\u5f15<\/span>\uff08domain Index\uff09<\/span>\u662f\u5df2\u77e5\u7684\u3002\u6211\u4eec\u7684\u7814\u7a76\u91cd\u662f\u5982\u4f55\u6709\u6548\u5229\u7528\u8fd9\u4e9b\u5206\u5e03\u53d8\u5316\u5b9e\u73b0\u6709\u6548\u7684\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u800c\u4e0d\u662f\u68c0\u6d4b\u5206\u5e03\u53d8\u5316\u3002\u5728LCD-NM [Yao er al., 2022]\u4e2d\uff0c\u4e3a\u4e86\u66f4\u7b80\u6d01\u5730\u63cf\u8ff0\u4e0d\u540cdomain\u4e4b\u95f4\u5206\u5e03\u7684\u533a\u522b\uff0c\u6211\u4eec\u5206\u522b\u5f15\u5165\u5728transition\u548cobservation function\u4e0a\u7684\u4f4e\u7ef4causal change factor \ud835\udec9\u3002\u5373\u5bf9\u4e8e\u4e0d\u540cdomain\u7684\u6570\u636e\uff0cchange factor \ud835\udec9\u53d6\u4e0d\u540c\u7684\u503c\u3002\u6700\u540e\uff0c\u6211\u4eec\u6269\u5c55\u4e0a\u8282\u4e2d\u63a2\u8ba8\u7684Nonparametric Latent Causal Processes\u6a21\u578b\u6765\u63cf\u8ff0\u6570\u636e\u5206\u5e03\u53d8\u5316\u4e0b\u7684\u6570\u636e\u751f\u6210\u8fc7\u7a0b\u3002\u901a\u8fc7\u7406\u8bba\u5206\u6790\uff0c\u6211\u4eec\u53d1\u73b0\u5728\u6a21\u578b\u4e2d\u5f15\u5165\u53d8\u5316\uff0c\u5e76\u5728\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u4f30\u8ba1\u6570\u636e\u7684\u53d8\u5316\uff0c\u56e0\u679c\u9690\u53d8\u91cf\u7684\u5b66\u4e60\u548c\u8bc6\u522b\u6548\u679c\u53ef\u4ee5\u5f97\u5230\u663e\u8457\u7684\u63d0\u5347\u3002<\/span><\/section>\n

<\/section>\n
\u4e0b\u9762\uff0c\u6211\u4eec\u5206\u522b\u4ee5\u89c6\u9891\u6570\u636e\u548c\u52a8\u4f5c\u6355\u6349\u6570\u636e\u4e3a\u7b97\u4f8b\uff0c\u9a8c\u8bc1\u6211\u4eec\u7684\u7406\u8bba\u548c\u6a21\u578b\u3002\u9996\u5148\uff0c\u6211\u4eec\u5728multi-domain cartpole\u6570\u636e\u4e0a\u6d4b\u8bd5\u3002\u5728\u8be5\u6570\u636e\u96c6\u7684\u4e0d\u540cdomain\u4e4b\u95f4\uff0c\u5c0f\u8f66\u7684\u8d28\u91cf\u548c\u91cd\u529b\u7cfb\u6570\u4f1a\u53d1\u751f\u53d8\u5316\uff0c\u4ece\u800c\u5bfc\u81f4\u65f6\u5e8f\u6570\u636e\u7684\u5206\u5e03\u53d1\u751f\u53d8\u5316\u3002\u6211\u4eec\u5728\u8be5\u6570\u636e\u96c6\u4e0a\u6d4b\u8bd5LCD-NM \uff0c\u53ef\u4ee5\u53d1\u73b0\u5728\u6570\u636e\u5206\u5e03\u53d1\u751f\u53d8\u5316\u7684\u60c5\u51b5\u4e0b\uff0c\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u5173\u7cfb\u53ef\u4ee5\u88ab\u8bc6\u522b\u3002<\/span><\/section>\n

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

<\/section>\n
\u540c\u65f6\uff0c\u6211\u4eec\u4e5f\u5728\u52a8\u4f5c\u6355\u6349\u6570\u636e<\/span>\uff08CMU-Mocap\uff09<\/span>\u4e0a\u9a8c\u8bc1\u7b97\u6cd5\u3002\u8be5\u7ec4\u6570\u636e<\/span>\uff08Subject #8\uff09<\/span>\u8bb0\u5f5511\u7ec4\u4e0d\u540c\u7c7b\u578b\u884c\u8d70<\/span>\uff08\u6bd4\u5982\uff0c\u5c0f\u788e\u6b65\u3001\u5927\u8de8\u6b65\uff0c\u5feb\u901f\u884c\u8d70\u7b49\uff09<\/span>\u8bd5\u9a8c\u6570\u636e\u3002\u8fd9\u4e9b\u4e0d\u540c\u7684\u884c\u8d70\u65b9\u5f0f\u5bfc\u81f4\u4e86\u6570\u636e\u5206\u5e03\u7684\u4e0d\u540c\u3002\u6211\u4eec\u53d1\u73b0LCD-NM\u53ef\u4ee5\u5f88\u597d\u7684\u5728\u65f6\u5e8f\u6570\u636e\u7684\u5206\u5e03\u53d1\u751f\u53d8\u5316\u7684\u60c5\u51b5\u4e0b\u6062\u590d\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u7684\u5173\u7cfb\u3002\u540c\u65f6\uff0c\u4f30\u8ba1\u7684\u56e0\u679c\u56e0\u5b50change factor\u4e5f\u80fd\u5f88\u597d\u7684\u63cf\u8ff0\u5206\u5e03\u7684\u533a\u522b\u3002<\/span><\/section>\n

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

<\/section>\n
\u6700\u540e\uff0c\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c[Yao et. al., 2021]\u4e2d\u63d0\u51fa\u7684nonstationary noise\u975e\u53c2\u6570\u6a21\u578b<\/span>\uff08LEAP\uff09<\/span>\u662fLCD-NM\u7684\u4e00\u4e2a\u7279\u4f8b\u3002\u76f8\u8f83\u800c\u8a00\uff0cLCD-NM\u8003\u8651\u4e86\u66f4\u666e\u9002\u7684\u5206\u5e03\u53d8\u5316\uff0c\u6bd4\u5982\u56e0\u679c\u5173\u7cfb\u7684\u53d8\u5316\uff0c\u540c\u65f6\u4e5f\u5229\u7528\u4e86nonstationary process\u4e2d\u90e8\u5206stationary process\u7684\u7279\u6027\u5b66\u4e60\u6765\u8bc6\u522b\u56e0\u679c\u9690\u53d8\u91cf\u3002<\/span><\/section>\n


<\/h2>\n

2.3 \u5206\u5e03\u8fc1\u79fb\u60c5\u5f62<\/span><\/strong><\/span><\/h2>\n

<\/section>\n
\u5728\u8bb8\u591a\u60c5\u51b5\u4e0b\uff0c\u8bb0\u5f55\u7684\u6570\u636e\u53ef\u80fd\u662f\u975e\u5e73\u7a33\u7684<\/span>\uff08non-stationary\uff09<\/span>\uff0c\u5176\u6570\u636e\u5206\u5e03\u968f\u65f6\u95f4\u800c\u53d8\u5316\uff0c\u6216\u8005\u662f\u5f02\u6784\u7684<\/span>\uff08heterogeneous\uff09<\/span>\uff0c\u5176\u6570\u636e\u5206\u5e03\u968f\u57df<\/span>\uff08domain\uff09<\/span>\u53d8\u5316\u3002\u8fd1\u6765\u7684\u7814\u7a76\u8868\u660e\uff0c\u6570\u636e\u5206\u5e03\u7684\u53d8\u5316\u53cd\u800c\u53ef\u4ee5\u7ed9\u56e0\u679c\u53d1\u73b0\u63d0\u4f9b\u66f4\u591a\u7684\u4fe1\u606f\uff0c\u4ece\u800c\u6709\u52a9\u4e8e\u56e0\u679c\u53d1\u73b0\u3002\u63a5\u4e0b\u6765\u7740\u91cd\u4ecb\u7ecd\u5728\u5206\u5e03\u53d1\u751f\u53d8\u5316\u7684\u65f6\u5019\uff0c\u4e0d\u540c\u60c5\u5883\u4e0b\u56e0\u679c\u53d1\u73b0\u7684\u51e0\u4e2a\u5de5\u4f5c\u3002<\/span><\/section>\n

<\/section>\n
Causal discovery from nonstationary\/heterogeneous data [CD-NOD, Huang et al., 2020]:<\/span><\/em><\/section>\n

<\/section>\n
CD-NOD \u7740\u529b\u4e8e\u89e3\u51b3\u5728\u5206\u5e03\u53d1\u751f\u53d8\u5316\u7684\u60c5\u51b5\u4e0b\uff0c\u5982\u4f55\u627e\u5230\u89c2\u6d4b\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\uff0c\u4e3b\u8981\u5305\u62ec\u4ee5\u4e0b\u4e09\u4e2a\u95ee\u9898\uff1a1. \u68c0\u6d4b\u54ea\u4e9b\u53d8\u91cf\u5bf9\u5e94\u7684\u56e0\u679c\u673a\u5236\u53d1\u751f\u4e86\u53d8\u5316\uff0c\u4ee5\u53ca\u5982\u4f55\u5b66\u4e60\u56e0\u679c\u9aa8\u67b6<\/span>\uff08causal skeleton\uff09<\/span>?  2. \u5982\u4f55\u5145\u5206\u5229\u7528\u6570\u636e\u5206\u5e03\u53d8\u5316\u643a\u5e26\u7684\u4fe1\u606f\u6765\u5224\u5b9a\u56e0\u679c\u65b9\u5411\uff1f3. \u627e\u5230\u4e86\u56e0\u679c\u7ed3\u6784\u4e4b\u540e\uff0c\u5982\u4f55\u8fdb\u4e00\u6b65\u5206\u6790\u56e0\u679c\u673a\u5236\u968f\u7740\u65f6\u95f4\u6216\u6761\u4ef6\u5177\u4f53\u662f\u5982\u4f55\u53d8\u5316\u7684\uff1f<\/span><\/section>\n

<\/section>\n
\u89c2\u6d4b\u53d8\u91cf\u5206\u5e03\u7684\u53d8\u5316\u53ef\u4ee5\u7406\u89e3\u4e3a\u5b58\u5728\u4e00\u4e9b\u672a\u89c2\u6d4b\u5230\u7684\u9690\u53d8\u91cf\u6216\u9690\u85cf\u7684\u6df7\u6742\u53d8\u91cf<\/span>\uff08hidden confounders\uff09<\/span>\uff0c\u5176\u503c\u53ef\u80fd\u968f\u7740\u65f6\u95f4\u6216\u57df\u53d8\u5316\uff0c\u4ece\u800c\u5bfc\u81f4\u89c2\u6d4b\u53d8\u91cf\u7684\u6982\u7387\u5206\u5e03\u53d1\u751f\u53d8\u5316\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u4e00\u4e2a\u53d8\u5316\u7684\u56e0\u679c\u5173\u7cfb\u662f\u56e0\u4e3a\u80cc\u540e\u5b58\u5728\u5f71\u54cd\u8fd9\u4e2a\u56e0\u679c\u8fc7\u7a0b\u7684\u53d8\u91cf\u3002\u4f8b\u5982\uff0c\u5728 fMRI \u6570\u636e\u4e2d\uff0c\u6570\u636e\u5206\u5e03\u53ef\u80fd\u4f1a\u968f\u7740\u6ce8\u610f\u529b\u6216\u5916\u90e8\u523a\u6fc0\u800c\u6539\u53d8\uff0c\u800c\u8fd9\u79cd\u523a\u6fc0\u53ef\u80fd\u4f1a\u6210\u4e3a\u5f71\u54cd\u5f88\u591a\u5927\u8111\u533a\u57df\u7684\u9690\u85cf\u6df7\u6742\u56e0\u7d20\u3002\u76f4\u63a5\u5ffd\u7565\u8fd9\u4e9b\u9690\u85cf\u7684\u6df7\u6742\u56e0\u7d20\u53ef\u80fd\u4f1a\u7ed9\u6211\u4eec\u5e26\u6765\u8bef\u5bfc\u6027\u7684\u7ed3\u679c\uff0c\u5305\u62ec\u865a\u5047\u7684\u56e0\u679c\u8fde\u63a5\u6216\u9519\u8bef\u7684\u56e0\u679c\u65b9\u5411\u3002<\/span><\/section>\n

<\/section>\n
CD-NOD\u7684\u57fa\u672c\u601d\u60f3\u5982\u4e0b\uff1a\u6211\u4eec\u5c06\u9690\u85cf\u7684\u53d8\u5316\u56e0\u7d20<\/span>\uff08hidden change factors\uff09<\/span>\u8868\u793a\u4e3aC\u7684\u51fd\u6570\uff0c\u5176\u4e2dC\u53ef\u4ee5\u662f\u57df\u7d22\u5f15<\/span>\uff08domain index\uff09<\/span>\uff0c\u65f6\u95f4\u7d22\u5f15<\/span>\uff08time index\uff09<\/span>\uff0c \u6216\u8005\u6761\u4ef6\u7d22\u5f15<\/span>\uff08condition index\uff09<\/span>\u7b49\u3002C \u5145\u5f53\u8868\u5f81\u6f5c\u5728\u53d8\u5316\u56e0\u7d20\u7684\u4ee3\u7406\u53d8\u91cf<\/span>\uff08surrogate variable\uff09<\/span>\u3002\u5728\u524d\u9762\u7684\u4f8b\u5b50\u4e2d\uff0c\u5927\u8111\u533a\u57df\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u53ef\u80fd\u4f1a\u88ab\u5e73\u6ed1\u53d8\u5316\u7684\u6ce8\u610f\u529b\u6216\u5916\u90e8\u523a\u6fc0\u6240\u6df7\u6dc6\uff0c\u56e0\u6b64\u8fd9\u91cc\u7684C\u53ef\u4ee5\u662f\u8868\u5f81\u6ce8\u610f\u529b\u53d8\u5316\u7684\u65f6\u95f4\u6307\u6807\u3002\u5728\u5904\u7406\u5305\u542b\u6765\u81ea\u4e0d\u540c\u53d7\u8bd5\u8005\u7684\u6570\u636e\u65f6\uff0c\u56e0\u679c\u5173\u7cfb\u53ef\u80fd\u4f1a\u53d7\u5230\u57fa\u56e0\u56e0\u7d20\u7684\u6df7\u6dc6\uff0c\u800c\u57fa\u56e0\u56e0\u7d20\u56e0\u53d7\u8bd5\u8005\u800c\u5f02\uff0c\u56e0\u6b64\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u53ef\u4ee5\u5c06C\u4f5c\u4e3a\u4e2a\u4f53\u7d22\u5f15<\/span>\uff08subject index\uff09<\/span>\uff0c\u4ee5\u6d88\u9664\u9690\u85cf\u57fa\u56e0\u56e0\u7d20\u7684\u5f71\u54cd\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u53d8\u91cf C \u7eb3\u5165\u56e0\u679c\u7cfb\u7edf\u4f5c\u4e3a\u8868\u5f81\u672a\u89c2\u5bdf\u5230\u7684\u53d8\u5316\u56e0\u7d20\u7684\u4ee3\u7406\u53d8\u91cf\uff0c\u7136\u540e\u5728\u589e\u5e7f\u53d8\u91cf\u96c6<\/span>\uff08augmented variable set\uff09<\/span>\u4e0a\u5e94\u7528\u57fa\u4e8e\u7ea6\u675f\u7684\u56e0\u679c\u53d1\u73b0\u7b97\u6cd5\u6765\u627e\u56e0\u679c\u7ed3\u6784\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5176\u4e2d\uff0cCD-NOD\u901a\u8fc7\u68c0\u6d4b\u89c2\u6d4b\u53d8\u91cfVi\u548cC\u4e4b\u95f4\u7684\u6761\u4ef6\u72ec\u7acb\u6027\uff0c\u6765\u68c0\u6d4b\u54ea\u4e9b\u53d8\u91cf\u5bf9\u5e94\u7684\u56e0\u679c\u673a\u5236\u53d1\u751f\u4e86\u53d8\u5316\u3002\u8fd9\u4e00\u6b65\u5728\u5f02\u5e38\u68c0\u6d4b\u548c\u6839\u672c\u539f\u56e0\u5206\u6790<\/span>\uff08root cause analysis\uff09<\/span>\u4e2d\u975e\u5e38\u6709\u7528\uff0c\u56e0\u4e3a\u5728\u4e00\u4e9b\u5b9e\u9645\u95ee\u9898\u4e2d\uff0c\u867d\u7136\u6211\u4eec\u53ef\u80fd\u4f1a\u89c2\u5bdf\u5230\u5927\u91cf\u53d8\u91cf\u7684\u8054\u5408\u5206\u5e03\u7684\u53d8\u5316\uff0c\u4f46\u5b9e\u9645\u4e0a\u53ea\u662f\u7531\u4e8e\u4e0e C \u6709\u76f4\u63a5\u8054\u7cfb\u7684\u5c11\u6570\u6839\u672c\u539f\u56e0\u3002<\/span><\/section>\n

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

\u771f\u5b9e\u7684\u56e0\u679c\u5173\u7cfb\u56fe\u3002\u5176\u4e2dg(C)\u548c\u03b8(C)\u8868\u793a\u672a\u89c2\u6d4b\u5230\u7684\u9690\u85cf\u7684\u53d8\u5316\u56e0\u7d20\uff0c\u5b83\u4eec\u53ef\u4ee5\u8868\u793a\u4e3a\u968f\u7740C\u53d8\u5316\u7684\u51fd\u6570\u3002<\/span><\/p>\n


<\/section>\n
\u8fdb\u4e00\u6b65\u5730\uff0c\u5728\u5b66\u4e60\u4e86\u56e0\u679c\u9aa8\u67b6\u4e4b\u540e\uff0c\u6211\u4eec\u901a\u8fc7\u72ec\u7acb\u53d8\u5316\u539f\u7406<\/span>\uff08independent change principle\uff09<\/span>\u6765\u5224\u5b9a\u56e0\u679c\u65b9\u5411\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u539f\u56e0\u7684\u5206\u5e03<\/span>\uff08P(cause)\uff09<\/span>\u548c\u7ed9\u5b9a\u539f\u56e0\u7684\u7ed3\u679c\u7684\u5206\u5e03<\/span>\uff08P(effect|cause)\uff09<\/span>\u662f\u72ec\u7acb\u53d8\u5316\u7684\uff0c\u4f46\u8fd9\u79cd\u72ec\u7acb\u6027\u901a\u5e38\u5728\u9519\u8bef\u7684\u65b9\u5411\u4e0a\u4e0d\u6210\u7acb\u3002\u5bf9\u4e8e\u4e24\u4e2a\u76f8\u90bb\u53d8\u91cf\uff0c\u5f53\u5176\u4e2d\u53ea\u6709\u4e00\u4e2a\u4e0e C \u76f4\u63a5\u8fde\u63a5\uff0c\u4e5f\u5c31\u662finvariance\u7684\u60c5\u51b5\uff0c\u662f\u4e00\u79cd\u7279\u6b8a\u60c5\u51b5\u3002<\/span><\/section>\n

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

CD-NOD\u641c\u7d22\u8fc7\u7a0b\u793a\u4f8b\u3002(a) \u521d\u59cb\u56fe\u3002\uff08b\uff09\u68c0\u6d4b\u54ea\u4e9b\u53d8\u91cf\u5bf9\u5e94\u7684\u56e0\u679c\u673a\u5236\u53d1\u751f\u4e86\u53d8\u5316\u3002\uff08c\uff09\u6062\u590d\u56e0\u679c\u9aa8\u67b6\u3002\uff08d\uff09\u8bc6\u522b\u56e0\u679c\u65b9\u5411\u3002<\/span><\/p>\n


<\/section>\n
\u4ece\u4e0a\u8ff0\u7684\u5206\u6790\u6211\u4eec\u5f88\u9ad8\u5174\u5730\u770b\u5230\u6982\u7387\u5206\u5e03\u7684\u53d8\u5316\u5176\u5b9e\u6709\u5229\u4e8e\u5b66\u4e60\u56e0\u679c\u7ed3\u6784\u7684\u3002\u56de\u60f3\u4e0b\uff0c\u5728\u72ec\u7acb\u540c\u5206\u5e03\u7684\u60c5\u51b5\u4e0b\uff0c\u5229\u7528\u566a\u58f0\u9879\u548c\u539f\u56e0\u4e4b\u95f4\u7684\u975e\u5bf9\u79f0\u72ec\u7acb\u6027\uff0c\u53ea\u6709\u5728\u9650\u5b9a\u7684\u51fd\u6570\u7a7a\u95f4\u4e0b\uff0c\u56e0\u679c\u65b9\u5411\u624d\u662f\u552f\u4e00\u786e\u5b9a\u7684\u3002\u6bd4\u5982\u8bf4\uff0c\u5728\u7ebf\u6027\u9ad8\u65af\u548c\u5728\u51fd\u6570\u7a7a\u95f4\u65e0\u4efb\u4f55\u9650\u5236\u7684\u60c5\u51b5\u4e0b\uff0c\u56e0\u679c\u65b9\u5411\u4e0d\u80fd\u552f\u4e00\u786e\u5b9a\u3002\u76f8\u53cd\u5730\uff0c\u5f53\u6709\u5206\u5e03\u53d8\u5316\u7684\u65f6\u5019\uff0c\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u72ec\u7acb\u53d8\u5316\u539f\u7406\u6765\u786e\u5b9a\u56e0\u679c\u65b9\u5411\uff0c\u800c\u4e0d\u9700\u8981\u9650\u5236\u56e0\u679c\u673a\u5236\u51fd\u6570\u7a7a\u95f4\u3002<\/span><\/section>\n

<\/section>\n

<\/section>\n
\u65f6\u53d8\u56e0\u679c\u6a21\u578b<\/span> (Time-varying causal model [Huang et al., 2019])<\/span>\uff1a<\/span><\/em><\/section>\n

<\/section>\n
\u5728\u8fd9\u4e2a\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u65f6\u53d8\u56e0\u679c\u6a21\u578b\uff0c\u6765\u8868\u793a\u975e\u5e73\u7a33\u65f6\u95f4\u5e8f\u5217\u4e2d\u7684\u56e0\u679c\u8fc7\u7a0b\u3002\u7279\u522b\u5730\uff0c\u6211\u4eec\u628a\u968f\u65f6\u95f4\u53d8\u5316\u7684\u56e0\u679c\u673a\u5236<\/span>\uff08causal mechanism) <\/span>\u770b\u505a\u9690\u53d8\u91cf\uff0c\u800c\u4e0d\u662f\u53c2\u6570\uff0c\u5e76\u4e14\u5bf9\u56e0\u679c\u673a\u5236\u7684\u53d8\u5316\u8fdb\u884c\u5efa\u6a21\u3002\u7279\u522b\u5730\uff0c\u5047\u8bbe\u53d8\u91cf\u4e4b\u95f4\u7684\u56e0\u679c\u5173\u7cfb\u662f\u7ebf\u6027\u7684\uff0c\u5e76\u4e14\u5176\u4e2d\u7684\u9690\u53d8\u91cf\uff1a\u56e0\u679c\u7cfb\u6570\u548c\u566a\u58f0\u9879\u7684\u65b9\u5dee\uff0c\u6ee1\u8db3auto-regressive models\u6765\u53d8\u5316\u3002\u5e76\u4e14\uff0c\u7531\u4e8e\u56e0\u679c\u6a21\u578b\u63cf\u8ff0\u4e86\u6982\u7387\u5206\u5e03\u662f\u5982\u4f55\u968f\u7740\u65f6\u95f4\u53d8\u5316\u7684\uff0c\u6240\u4ee5\u5982\u679c\u6211\u4eec\u77e5\u9053\u56e0\u679c\u6a21\u578b\u662f\u5982\u4f55\u968f\u7740\u65f6\u95f4\u53d8\u5316\u7684\uff0c\u5c31\u53ef\u4ee5\u4ee5\u6b64\u505a\u51fa\u9002\u5e94\u6027\u9884\u6d4b\u3002<\/span><\/section>\n

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

Figure: \u65f6\u53d8\u56e0\u679c\u6a21\u578b\u793a\u610f\u56fe\u3002\u5176\u4e2db\u548ch\u662f\u9690\u53d8\u91cf\uff0c\u4ed6\u4eec\u5206\u522b\u7528\u6765\u8868\u5f81\u6ee1\u8db3auto-regressive model\u7684\u56e0\u679c\u7cfb\u6570\u548c\u566a\u58f0\u9879\u65b9\u5dee\u7684\u5bf9\u6570\u3002<\/span><\/p>\n


<\/section>\n
\u4e4b\u540e\u7684\u8bc6\u522b\u6f5c\u5728\u7684\u56e0\u679c\u8fc7\u7a0b<\/span>\uff08identifying latent causal process\uff09<\/span>\u7684\u5de5\u4f5c\uff0c\u6bd4\u5982 [Yao et al., 2021\uff0c2022]\uff0c\u53ef\u4ee5\u770b\u505a\u66f4\u8fdb\u4e00\u6b65\u7684\u6269\u5c55\u3002\u5176\u5141\u8bb8\u8f93\u5165\u53d8\u91cf\u662f\u56fe\u50cf\uff0c\u5e76\u4e14\u56e0\u679c\u5173\u7cfb\u53ef\u4ee5\u662f\u975e\u7ebf\u6027\u7684\u3002<\/span><\/section>\n

<\/section>\n

<\/section>\n
Partial Disentanglement for Domain Adaption [Kong et. al., 2022]<\/span><\/em><\/section>\n
<\/span><\/section>\n
 <\/span><\/section>\n
\u56fe\u7247\u8bc6\u522b\u7684\u8fc1\u79fb\u5b66\u4e60<\/span>\uff08Domain Adaptation<\/span>\uff09\u4efb\u52a1\u901a\u5e38\u4f1a\u63d0\u4f9b\u591a\u7ec4\u5728\u4e0d\u540csource domain\u4e0b\u56fe\u7247x\u548c\u6807\u7b7ey\u3002\u5176\u6761\u4ef6\u5206\u5e03P(y|x)\u5728\u4e0d\u540cdomain\u4e0b\u6709\u4e00\u5b9a\u53d8\u5316\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u5728\u4f7f\u5728source domain\u4e0a\u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u901a\u8fc7\u8fc1\u79fb\u8bad\u7ec3\uff0c\u80fd\u5728\u672a\u63d0\u4f9b\u6807\u7b7e\u7684target domain\u4e0b\u5177\u6709\u826f\u597d\u8868\u73b0\u3002\u4f20\u7edf\u8fc1\u79fb\u5b66\u4e60\u7740\u91cd\u4e8e\u5b66\u4e60\u4e0d\u540cdomain\u4e4b\u95f4\u4e0d\u53d8\u7684\u90e8\u5206\uff0c\u5e76\u4f7f\u7528\u8fd9\u4e9b\u4e0d\u53d8\u7684\u90e8\u5206\u8fdb\u884c\u9884\u6d4b\u3002\u4e0e\u4e4b\u4e0d\u540c\u7684\u662f\uff0c\u6211\u4eec\u63d0\u51fa\u7684iMSDA [Kong et. al., 2022]\u901a\u8fc7\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u8bc6\u522b\u51fa\u89e3\u91cadomain\u4e4b\u95f4\u533a\u522b\u7684\u56e0\u679c\u9690\u53d8\u91cf\uff0c\u5e76\u4f7f\u7528\u8fd9\u4e9b\u9690\u53d8\u91cf\u5bf9\u6570\u636e\u505atranslation\uff0c\u5c06\u5176\u6295\u5c04\u5230\u540c\u4e00domain\u8fdb\u884c\u8bad\u7ec3\uff0c\u5b9e\u73b0\u8fc1\u79fb\u5b66\u4e60\u3002<\/span><\/section>\n

<\/span><\/section>\n\n\n\n
\n
<\/section>\n<\/td>\n
\n
<\/section>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

<\/section>\n
\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u6211\u4eec\u5047\u8bbe\u56fe\u7247x\u548c\u6807\u7b7ey\u662f\u7531\u56e0\u679c\u9690\u53d8\u91cfzc<\/sub>\u548czs<\/sub>\u901a\u8fc7\u975e\u7ebf\u6027<\/span>\uff08\u4f46\u53ef\u9006\uff09<\/span>\u7684\u6620\u5c04\u751f\u6210\u3002\u5176\u4e2d\uff0czc<\/sub>\u5305\u542b\u56fe\u7247content\u4fe1\u606f<\/span>\uff08domain\u4e4b\u95f4\u4e0d\u53d8\u7684\u90e8\u5206\uff09<\/span>\uff0czs<\/sub>\u5305\u542b\u56fe\u7247style\u4fe1\u606f<\/span>\uff08domain\u4e4b\u95f4\u53d8\u7684\u90e8\u5206\uff09<\/span>\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u4e0d\u540cdomain\u4e0b\uff0c\u751f\u6210\u51fd\u6570\u662f\u4e0d\u53d8\u7684\uff0c\u53d8\u5316\u7684\u90e8\u5206\u4ec5\u6709zs<\/sub>\u90e8\u5206\u7684\u5206\u5e03\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u5b66\u4e60\u751f\u6210\u51fd\u6570\u5176\u5b9e\u662f\u6240\u6709\u56fe\u7247\u7684\u5171\u6709\u7684\u6e32\u67d3\u7a0b\u5e8f\u3002\u4e3a\u4e86\u56e0\u679c\u9690\u53d8\u91cfzs<\/sub>\u7684\u53ef\u8bc6\u522b\u6027\uff0c\u6211\u4eec\u5047\u8bbezs<\/sub>\u5728\u7ed9\u5b9adomain index u\u7684\u60c5\u51b5\u4e0b\u7684\u6761\u4ef6\u72ec\u7acb\u6027<\/span>\uff08\u8fd9\u662f\u4e00\u79cd\u5b9e\u73b0minimal change principle\u7684\u65b9\u5f0f\uff09<\/span>\u3002\u901a\u8fc7\u5b66\u4e60\u5e76\u8bc6\u522bzs<\/sub>\u548c\u5176\u751f\u6210\u8fc7\u7a0b\uff0c\u6211\u4eec\u53ef\u4ee5multi-domain\u7684\u56fe\u7247\uff0c\u53ef\u8bc6\u522b\u5730translate\u5230target domain\uff0c\u5982\u4e0a\u56fe\u5c06\u6240\u6709\u989c\u8272\u6570\u636e\u90fd\u6295\u5c04\u5230\u7ea2\u8272\u6570\u5b57domain\u4e2d\uff0c\u4ece\u800c\u76f4\u63a5\u5728target domain<\/span>\uff08\u7ea2\u8272\u6570\u5b57\uff09<\/span>\u4e0a\u89e3\u51b3\u8fc1\u79fb\u5b66\u4e60\u95ee\u9898\u3002<\/span><\/section>\n

<\/section>\n
\u6700\u540e\uff0c\u6211\u4eec\u7edf\u4e00\u63a2\u8ba8\u72ec\u7acb\u6210\u5206\u5206\u6790<\/span>\uff08ICA\uff09<\/span>\u548c\u56e0\u679c\u5b66\u4e60\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u56e0\u679c\u53d1\u73b0\u662fICA\u95ee\u9898\u7684\u4e00\u79cd\u7279\u6b8a\u60c5\u51b5\u3002ICA\u89e3\u51b3\u7684\u95ee\u9898\u662f\u4ece<\/span>\uff08\u7ebf\u6027\u6216\u975e\u7ebf\u6027\uff09<\/span>\u6df7\u6742\u7684\u4fe1\u53f7\u4e2d\u8bc6\u522b\u51fa\u76f8\u4e92\u72ec\u7acb\u7684\u6210\u5206\u3002\u800c\u51fd\u6570\u529f\u80fd\u56e0\u679c\u6a21\u578b<\/span>\uff08FCM\uff09<\/span>\u7684\u6570\u636e\u751f\u6210\u8fc7\u7a0b\u6b63\u662f\u5c06\u76f8\u4e92\u72ec\u7acb\u7684\u566a\u58f0\uff0c\u901a\u8fc7\u56e0\u679c\u5173\u7cfb\u6df7\u6742\u6210\u89c2\u6d4b\u5230\u6570\u636e\u3002\u4e24\u8005\u552f\u4e00\u7684\u533a\u522b\u662f\uff0cFCM\u5728\u7ed3\u6784\u4e0a\u9700\u6ee1\u8db3\u6709\u5411\u975e\u73af\u56fe<\/span>\uff08DAG\uff09<\/span>\uff0c\u6216\u5728\u65f6\u5e8f\u6570\u636e\u4e0a\u6ee1\u8db3Granger Causality\u5047\u8bbe\u7b49\u3002\u7ebf\u6027ICA\u7684\u53ef\u8bc6\u522b\u6027\u53ea\u9700\u6ee1\u8db3non-Gaussian source\u5c31\u53ef\u8fbe\u5230\uff0c\u56e0\u6b64LiNGAM\u80fd\u5c06ICA\u76f4\u63a5\u5e94\u7528\u5230\u7ebf\u6027\u6761\u4ef6\u4e0b\u7684\u56e0\u679c\u53d1\u73b0\u95ee\u9898\u3002\u7136\u800c\uff0cnonlinear ICA\u7684\u53ef\u8bc6\u522b\u6027\u662f\u975e\u5e38\u56f0\u96be\u7684\u95ee\u9898\u3002\u76f4\u5230\u8fd1\u4e24\u5e74\uff0c\u4eba\u4eec\u624d\u53d1\u73b0\u5728\u80fd\u627e\u5230\u591a\u4e2adomain\u7684\u6570\u636e\uff0c\u5e76\u4e14\u56e0\u679c\u9690\u53d8\u91cf\u5728\u989d\u5916\u63d0\u4f9b\u7684surrogate variable\u4e0b\u88ab\u6761\u4ef6\u72ec\u7acb\u65f6\uff0c\u6211\u4eec\u624d\u80fd\u6062\u590d\u51fa\u8fd9\u4e9b\u56e0\u679c\u9690\u53d8\u91cf\u3002\u5728[Ricardo Pio et al. 2022] \u4e2d\uff0c\u6211\u4eec\u5c31\u662f\u4f7f\u7528\u8fd9\u79cdnonlinear ICA\u7684\u53ef\u8bc6\u522b\u6027\uff0c\u5b9e\u73b0\u975e\u7ebf\u6027\u7684\u56e0\u679c\u53d1\u73b0\u7684\u3002<\/span><\/section>\n

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

<\/section>\n
\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u5219\u66f4\u662f\u4e0eICA\u606f\u606f\u76f8\u5173\u3002\u4f8b\u5982\uff0c\u4e4b\u524d\u6211\u4eec\u8ba8\u8bba\u7684\u65f6\u5e8f\u56e0\u679c\u8868\u5f81\u5b66\u4e60[Yao et al., 2021 & 2022]\u548c\u8fc1\u79fb\u5b66\u4e60[Kong et. al., 2022]\uff0c\u90fd\u53ef\u89c6\u4e3a\u8fd9\u7c7b\u60c5\u51b5\u4e0b\u7684nonlinear ICA\u5728\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u9886\u57df\u7684\u62d3\u5c55\u3002\u7136\u800c\uff0c\u8fd9\u7c7bICA\u65b9\u6cd5\u7684\u9650\u5236\u5728\u4e8e\uff0c\u8bad\u7ec3\u6570\u636e\u5fc5\u987b\u542b\u6709\u8db3\u4ee5\u89e3\u91cadomain\u533a\u522b\u7684surrogate variable\uff0c\u800c\u4e14\u8fd9\u4e9bsurrogate variable\u53ef\u4f7f\u56e0\u679c\u9690\u53d8\u91cf\u72ec\u7acb\u3002\u81ea\u7136\uff0c\u6211\u4eec\u4f1a\u6709\u7591\u95ee\uff1a\u5728\u6ca1\u6709surrogate variable\u7684\u60c5\u51b5\u4e0b\uff0cnonlinear ICA\u662f\u5426\u4e5f\u80fd\u8fbe\u5230\u7684\u53ef\u8bc6\u522b\u6027\u5462\uff1f\u5728[Zheng et. al., 2022]\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528structural sparsity assumption\u6765\u56de\u7b54\u8fd9\u4e2a\u95ee\u9898\u3002\u6211\u4eec\u53d1\u73b0\u5982\u679c\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\uff0c\u5bf9\u6bcf\u4e2a\u9690\u53d8\u91cfSi<\/sub>\uff0c\u6211\u4eec\u90fd\u80fd\u627e\u5230\u4e00\u7ec4\u53d8\u91cf{Xi<\/sub>}\uff0c\u4f7f\u5f97Si<\/sub>\u662f\u8fd9\u7ec4\u53d8\u91cf\u552f\u4e00\u7684\u5171\u6709parent\u65f6\uff0cnonlinear ICA\u5c31\u53ef\u4ee5\u8fbe\u5230\u53ef\u8bc6\u522b\u6027\u3002\u4f8b\u5982\uff0c\u5728\u4e0a\u56fe\u4e2d\uff0c\u5bf9\u4e8eX1\u548cX4\uff0cS1\u662f\u5b83\u4eec\u4e4b\u95f4\u552f\u4e00\u5171\u6709\u7684parent\u3002\u5728\u8fd9\u79cd\u7ed3\u6784\u7a00\u758f\u6027\u7684\u6761\u4ef6\u4e0b\uff0c\u6211\u4eec\u53ef\u4ee5\u5728\u6ca1\u6709surrogate variable\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0nonlinear ICA\u7684\u53ef\u8bc6\u522b\u6027\u3002\u8fd9\u7c7bsparse nonlinear ICA\u5728\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u4e0a\u7684\u62d3\u5c55\u4f1a\u662f\u4e00\u4e2a\u5f88\u6709\u610f\u601d\u7684\u7814\u7a76\u65b9\u5411\u3002<\/span><\/section>\n

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


<\/p>\n

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

\n
\n
\n

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

\n
\n
\n
\n
Spirtes, P., Meek, C., and Richardson, T. Causal inference in the presence of latent variables and selection bias. In Proceedings of the Eleventh conference on Uncertainty in artificial intelligence (UAI 1995), pp. 499\u2013506, 1995.<\/span><\/section>\n

<\/section>\n
Peter Spirtes, Clark Glymour, Richard Scheines. Causation, prediction, and search [M]. London: MIT press, 2000.<\/span><\/section>\n

<\/section>\n
Charles Spearman. Pearson\u2019s contribution to the theory of two factors [J]. British Journal of Psychology, 1928, 19(1):95\u2013101.<\/span><\/section>\n

<\/section>\n
Ricardo Silva, Richard Scheines, Clark Glymour, and Peter Spirtes. Learning the structure of linear latent variable models [J]. Journal of Machine Learning Research, 2006, 7(Feb): 191-246.<\/span><\/section>\n

<\/section>\n
Ricardo Silva, Richard Scheines, Clark Glymour, and Peter Spirtes. Learning the structure of linear latent variable models [J]. Journal of Machine Learning Research, 2006, 7(Feb): 191-246.<\/span><\/section>\n

<\/section>\n
Erich Kummerfeld, and Joseph Ramsey. Causal clustering for 1-factor measurement models[C]. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD 2016). New York: ACM, 2016: 1655-1664.<\/span><\/section>\n

<\/section>\n
H. Cram\u00e9r. Random variables and probability distributions (2nd edition) [M]. Cambridge: Cambridge University Press, 1962.<\/span><\/section>\n

<\/section>\n
Peter Spirtes, and Kun Zhang. Causal discovery and inference: concepts and recent methodological advances [J]. Applied informatics, 2016, 3(1): 1-28.<\/span><\/section>\n

<\/section>\n
F. Xie, R. Cai, B. Huang, C. Glymour, Z. Hao, and K. Zhang. Generalized Independent Noise Condition for Estimating Latent Variable Causal Graphs. Advances in Neural Information Processing Systems 33 (NeurIPS 2020), Virtual Conference, 2020.<\/span><\/section>\n

<\/section>\n
R. Cai, F. Xie, C. Glymour, Z. Hao, and K. Zhang. Triad Constraints for Causal Discovery in the Presence of Latent Variables. Advances in Neural Information Processing Systems 32 (NeurIPS 2019), Vancouver, CANADA, 2019.<\/span><\/section>\n

<\/section>\n
Z. Chen, F. Xie, J. Qiao, Z. Hao, K. Zhang, and R. Cai. Identification of Linear Latent Variable Model with Arbitrary Distribution. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI 2022), Vancouver, CANADA, 2022.<\/span><\/section>\n

<\/section>\n
F. Xie, B. Huang, Z. Chen, Y. He, Z. Geng, and K. Zhang. Identification of Linear Non-Gaussian Latent Hierarchical Structure. Thirty-ninth International Conference on Machine Learning (ICML 2022), Baltimore, Maryland USA, 2022.<\/span><\/section>\n

<\/section>\n
Adams, J., Hansen, N., and Zhang, K. Identification of partially observed linear causal models: Graphical conditions for the non-gaussian and heterogeneous cases. Advances in Neural Information Processing Systems (NeurIPS 2019), 34, 2021.<\/span><\/section>\n

<\/section>\n
Zhang, K. and Hyv\u00a8arinen, A. On the identifiability of the post-nonlinear causal model. In Proceedings ofthe twenty-fifth conference on uncertainty in artificial intelli- gence, pp. 647\u2013655. AUAI Press, 2009.<\/span><\/section>\n

<\/section>\n
W. Yao, Y. Sun, A. Ho, C. Sun and K. Zhang. Learning Temporally Causal Latent Processes from General Temporal Data. ICLR 2022.<\/span><\/section>\n

<\/section>\n
W. Yao, G. Chen and K. Zhang. Causal Disentanglement for Time Series. NeurIPS 2022.<\/span><\/section>\n

<\/section>\n
Y. Zheng, I. Ng and K. Zhang. On the Identifiability of Nonlinear ICA: Sparsity and Beyond. <\/span>arXiv preprint arXiv:2206.07751<\/span> (2022). <\/span><\/span><\/section>\n

<\/section>\n
Monti, Ricardo Pio, Kun Zhang, and Aapo Hyv\u00e4rinen. “Causal discovery with general non-linear relationships using non-linear ica.” <\/span>Uncertainty in artificial intelligence<\/span>. PMLR, 2020.<\/span><\/span><\/section>\n

Kong, Lingjing, et al. “Partial disentanglement for domain adaptation.” <\/span>International Conference on Machine Learning<\/span>. PMLR, 2022.<\/span><\/span><\/section>\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


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

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

\n
\n
\n
\n
\n

<\/section>\n

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

3. \u5e94\u7528<\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

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


<\/p>\n

\u5728\u672c\u7ae0\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u5728\u8fc1\u79fb\u5b66\u4e60\u3001\u9002\u5e94\u6027\u5f3a\u5316\u5b66\u4e60\u3001\u4ee5\u53ca\u52a8\u4f5c\u8bc6\u522b\u7b49\u673a\u5668\u5b66\u4e60\u4efb\u52a1\u4e0a\u7684\u5e94\u7528\u3002<\/span><\/section>\n

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

3.1 \u8fc1\u79fb\u5b66\u4e60<\/span><\/strong><\/span><\/h2>\n

<\/section>\n
\u8fc1\u79fb\u5b66\u4e60\u662f\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u76f4\u63a5\u7684\u5e94\u7528\u4e4b\u4e00\u3002\u6839\u636eMinimal Change Principle [Huang et al., 2020]\uff0c\u5728\u6b63\u786e\u7684\u56e0\u679c\u8868\u5f81\u4e0b\uff0c\u4e0d\u540cdomain\u4e4b\u95f4\u7684\u5206\u5e03\u53d8\u5316\u7684\u901a\u5e38\u53ef\u7531\u5728\u5c11\u6570\u56e0\u679c<\/span>\uff08\u9690\uff09<\/span>\u53d8\u91cf\u4e4b\u95f4\u7684causal edge\/node\u4e0a\u7684\u5e72\u9884\u6765\u89e3\u91ca\u3002\u56e0\u6b64\uff0c\u5982\u679c\u56e0\u679c\u9690\u53d8\u91cf\u548c\u5b83\u4eec\u4e4b\u95f4\u7684\u5173\u7cfb\u80fd\u88ab\u8bc6\u522b\uff0c\u8fc1\u79fb\u5b66\u4e60\u5c31\u53ea\u7528\u53d1\u751f\u5728\u6781\u5c11\u6570\u7684\u6a21\u578b\u53c2\u6570\u4e0a\u3002\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u5927\u5927\u51cf\u5c11\u4e86\u8fc1\u79fb\u5b66\u4e60\u9700\u8981\u7684\u6570\u636e\u91cf\u3002\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u4eba\u7c7b\u5728\u65b0\u73af\u5883\u5728\u7075\u6d3b\u53d8\u5316\u3001\u9002\u5e94\u7684\u6839\u672c\u539f\u56e0\u3002\u5728[Zhang et al., 2020]\u4e2d\uff0c\u6211\u4eec\u9996\u5148\u4f7f\u7528\u4e0a\u7ae0\u8282\u4ecb\u7ecd\u7684\uff0c\u5728\u6570\u636e\u5206\u5e03\u53d1\u751f\u53d8\u5316\u7684\u60c5\u51b5\u4e0b\u56e0\u679c\u53d1\u73b0\u65b9\u6cd5CD-NOD [Huang et. al., 2020]\uff0c\u5bfb\u627e(X\uff0cY)\u4e4b\u95f4\u7684\u56e0\u679c\u56fe\u3002\u5982\u53f3\u56fe\u6240\u793a\uff0c\u5728\u627e\u51fa\u6b63\u786e\u7684causal factorization\u540e\uff0c\u6211\u4eec\u53ef\u5bf9\u56fe\u91cc\u7684\u6bcf\u7ec4\u72ec\u7acb\u56e0\u679c\u6a21\u5757\uff0c\u6bd4\u5982P(X3<\/sub> | X2<\/sub>, Y, \ud835\udec93<\/sub>)\uff0c\u5355\u72ec\u5efa\u6a21\uff0c\u901a\u8fc7\u56e0\u679c\u53d8\u5316\u56e0\u5b50\ud835\udec9\u6355\u6349\u4e0d\u540csource domain\u4e4b\u95f4\u7684\u5206\u5e03\u533a\u522b\u3002\u5728\u505a\u8fc1\u79fb\u65f6\uff0c\u6211\u4eec\u53ea\u9700\u4f7f\u7528\u5c11\u91cftarget domain\u6570\u636e\uff0c\u63a8\u65ad\u6216\u5b66\u4e60\u51fa\u53d1\u751f\u53d8\u5316\u7684causal edge\u4ee5\u53ca\u8be5edge\u4e0a\u7684\u53d8\u5316\u56e0\u5b50\uff0c\u5c31\u53ef\u5b9e\u73b0\u6709\u6548\u7684\u8fc1\u79fb\u5b66\u4e60\u3002\u5f53\u89c2\u6d4b\u5230\u7684\u6570\u636e<\/span>\uff08\u6bd4\u5982\u56fe\u50cf\u50cf\u7d20\uff09<\/span>\u4e0d\u662f\u56e0\u679c\u76f8\u5173\u7684\u53d8\u91cf\u800c\u662f\u7531\u56e0\u679c\u9690\u53d8\u91cf\u751f\u6210\u662f\uff0c\u6211\u4eec\u53ef\u4f7f\u7528\u4e0a\u8282\u63a2\u8ba8\u7684iMSDA [Kong et. al., 2022]\uff0c\u901a\u8fc7\u56e0\u679c\u8868\u5f81\u5b66\u4e60\uff0c\u8bc6\u522b\u51fa\u89e3\u91cadomain\u4e4b\u95f4\u53d8\u5316\u7684\u56e0\u679c\u9690\u53d8\u91cf\uff0c\u5e76\u4f7f\u7528\u8fd9\u4e9b\u9690\u53d8\u91cf\u5bf9\u56fe\u50cf\u6570\u636e\u505atranslation\uff0c\u5c06\u5176\u6295\u5c04\u5230\u540c\u4e00domain\u8fdb\u884c\u8bad\u7ec3\uff0c\u4ece\u800c\u5b9e\u73b0\u56fe\u50cf\u8fc1\u79fb\u5b66\u4e60\u3002<\/span><\/section>\n

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

<\/section>\n
\u53e6\u5916\uff0c\u57fa\u4e8eMinimal Change Principle\u7684\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u4e5f\u53ef\u5e94\u7528\u4e8e\u56fe\u7247\u8f6c\u6362<\/span>\uff08Image Translation\uff09<\/span>\u4efb\u52a1\u3002\u5728\u8bad\u7ec3\u6570\u636e\u96c6\u5305\u542b\u591a\u4e2a\u672a\u5339\u914d\u7684\u56fe\u50cfdomain\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u63d0\u51fa\u7684\u57fa\u4e8e\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u7684\u6a21\u578bIrwGAN [Xie et. al., 2021]\u53ef\u4ee5\u6709\u6548\u5730\u5bf9\u4e0d\u540cdomain\u7684\u56fe\u7247\u8fdb\u884c\u5339\u914d\u540c\u65f6\u5b66\u4e60\u56fe\u7247\u8f6c\u6362\u51fd\u6570\uff0c\u89e3\u51b3\u56fe\u7247\u8f6c\u6362\u8fc7\u7a0b\u4e2dcontent\u548cstyle\u6df7\u6742\u7684\u95ee\u9898\u3002<\/span><\/section>\n

<\/span><\/section>\n
<\/section>\n
 <\/span><\/section>\n
\u5982\u4e0a\u56fe\u6240\u793a\uff0c\u4e0e\u5176\u4f59\u73b0\u6709\u65b9\u6cd5\u76f8\u6bd4\uff0c\u57fa\u4e8e\u56e0\u679c\u8868\u5f81\u5b66\u4e60\u7684\u56fe\u50cf\u8f6c\u6362\u65b9\u6cd5IrwGAN\u4fdd\u7559\u4e86\u4e3b\u8981\u7684\u56fe\u7247\u5185\u5bb9\u4fe1\u606f\uff0c\u5e76\u4e14\u4ec5\u5bf9\u56fe\u7247\u7684\u98ce\u683c\u4fe1\u606f\u8fdb\u884c\u7f16\u8f91\u548c\u8f6c\u6362\u3002<\/span><\/section>\n


<\/h2>\n

3.2 \u9002\u5e94\u6027\u5f3a\u5316\u5b66\u4e60<\/span><\/strong><\/span><\/h2>\n

<\/section>\n
\u5f3a\u5316\u5b66\u4e60\u662f\u901a\u8fc7\u4e0e\u5916\u754c\u73af\u5883\u4e92\u52a8\u6765\u5b66\u4e60\u7b56\u7565\u4f7f\u5f97\u5956\u52b1\u6700\u5927\u5316\u3002\u5f3a\u5316\u5b66\u4e60\u7684\u4e00\u4e2a\u5b9e\u9645\u6311\u6218\u662f\u5982\u4f55\u5728\u9762\u5bf9\u65b0\u73af\u5883\u65f6\u5bf9\u4e4b\u524d\u5b66\u4e60\u5230\u7684\u7b56\u7565\u8fdb\u884c\u5feb\u901f\u4e14\u6709\u6548\u7684\u8c03\u6574\uff0c\u56e0\u4e3a\u73b0\u5b9e\u4e16\u754c\u603b\u662f\u5728\u53d8\u5316\u7684\u3002<\/span><\/section>\n

<\/section>\n
\u9ec4\u7b49\u63d0\u51fa\u7684AdaRL [Huang et al., 2022]\u53ef\u4ee5\u6709\u6548\u5730\u9002\u5e94\u8de8\u9886\u57df\u6700\u4f18\u7b56\u7565\u7684\u53d8\u5316\u3002\u5176\u57fa\u672c\u601d\u60f3\u662f\uff1a\u627e\u5230\u5f71\u54cd\u6700\u4f18\u7b56\u7565\u7684\u6700\u5c0f\u5145\u5206\u56e0\u7d20\u96c6<\/span>\uff08a minimal sufficient set of factors\uff09<\/span>\uff0c\u5e76\u7406\u89e3\u8fd9\u4e9b\u56e0\u7d20\u5728\u4e0d\u540c\u73af\u5883\u4e2d\u662f\u5982\u4f55\u53d8\u5316\u7684\u3002\u4f8b\u5982\uff0c\u5728Cartpole\u6e38\u620f\u4e2d\uff0c\u5f53\u6211\u4eec\u5c06\u7cfb\u7edf\u4ece\u5730\u7403\u79fb\u52a8\u5230\u706b\u661f\u65f6\uff0c\u5982\u679cagent\u53ef\u4ee5\u523b\u753b\u662f\u91cd\u529b\u6539\u53d8\u4e86\u6700\u4f18\u7b56\u7565\uff0c\u5e76\u7406\u89e3\u5b83\u5982\u4f55\u5f71\u54cd\u7b56\u7565\uff0c\u90a3\u4e48agent\u5c31\u53ef\u4ee5\u901a\u8fc7\u4ec5\u7528\u5c11\u91cf\u6837\u672c\u4f30\u8ba1\u4e00\u7ef4\u91cd\u529b\u6765\u5feb\u901f\u9002\u5e94\u7b56\u7565\u3002\u4e0e\u76f4\u63a5\u9002\u5e94\u89c2\u5bdf\u5230\u7684\u9ad8\u7ef4\u56fe\u50cf\u76f8\u6bd4\uff0c\u5b83\u66f4\u6709\u6548\uff0c\u4e5f\u66f4\u6613\u4e8e\u89e3\u91ca\u3002<\/span><\/section>\n

<\/section>\n
\u73b0\u5728\u8ba9\u6211\u4eec\u8be6\u7ec6\u770b\u4e00\u4e0b\u56fe\u8868\u793a<\/span> (graphical representation) <\/span>\u548c\u6a21\u5757\u5316\u5c5e\u6027<\/span> (modularity)<\/span> \u662f\u5982\u4f55\u5e2e\u52a9\u8fc1\u79fb\u5f3a\u5316\u5b66\u4e60\u3002<\/span><\/section>\n

<\/section>\n
\u5047\u8bbe\u6211\u4eec\u6709\u6765\u81ea\u591a\u4e2a\u9886\u57df\u7684\u6570\u636e\uff0c\u5305\u62ec\u9ad8\u7ef4\u56fe\u50cf\u5e8f\u5217\u3001\u52a8\u4f5c\u548c\u5956\u52b1\uff0c\u5e76\u4e14\u4e0d\u540c\u7684\u9886\u57df\u53ef\u80fd\u6709\u4e0d\u540c\u7684\u73af\u5883\uff0c\u6bd4\u5982\u91cd\u529b\u548c\u56fe\u50cf\u4e0a\u6240\u52a0\u566a\u58f0\u6c34\u5e73\u7684\u53d8\u5316\u3002AdaRL\u7684\u57fa\u672c\u601d\u60f3\u662f\u6211\u4eec\u5229\u7528\u7b80\u7ea6\u7684\u56fe\u8868\u793a<\/span>\uff08parsimonious graphical representation\uff09<\/span>\u6765\u8868\u5f81\u7cfb\u7edf\u4e2d\u53d8\u91cf\u4e4b\u95f4\u7684\u7ed3\u6784\u5173\u7cfb\uff0c\u5305\u62ec\u4e0d\u540c\u7ef4\u5ea6\u7684\u72b6\u6001s\u3001\u4ee5\u53ca\u5956\u52b1\u53d8\u91cfr\u3001\u52a8\u4f5c\u53d8\u91cfa\u3001\u4ee5\u53ca\u611f\u77e5\u56fe\u50cfo\u4e4b\u95f4\u7684\u7ed3\u6784\u3002\u6b64\u5916\uff0c\u4e0e\u5b66\u4e60\u9ad8\u7ef4\u8f93\u5165\u4e0a\u7684\u5206\u5e03\u504f\u79fb\u7684\u7ecf\u5178\u65b9\u6cd5\u4e0d\u540c\uff0c\u7531\u4e8e\u56fe\u5206\u89e3\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4f4e\u7ef4\u5411\u91cf \u03b8 \u4ee5\u7d27\u51d1\u7684\u65b9\u5f0f\u8868\u5f81\u7279\u5b9a\u9886\u57df\u7684\u4fe1\u606f\u3002\u7279\u522b\u5730\uff0c\u03b8o<\/sup> \u7528\u4e8e\u8868\u5f81\u89c2\u5bdf\u51fd\u6570\u7684\u53d8\u5316\uff0c\u03b8s <\/sup>\u7528\u4e8e\u8f6c\u6362\u52a8\u6001\u7684\u53d8\u5316\uff0c\u03b8r<\/sup> \u7528\u4e8e\u5956\u52b1\u51fd\u6570\u7684\u53d8\u5316\u3002\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u57df\u5728\u91cd\u529b\u548c\u6240\u52a0\u767d\u566a\u58f0\u65b9\u9762\u662f\u4e0d\u540c\u7684\uff0c\u56e0\u6b64\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u4e00\u4e2a\u4e8c\u7ef4\u5411\u91cf\u6765\u8868\u5f81\u8fd9\u4e9b\u53d8\u5316\uff0c\u5176\u4e2d\u4e00\u7ef4\u7528\u4e8e\u91cd\u529b\uff0c\u53e6\u4e00\u7ef4\u7528\u4e8e\u4e0d\u540c\u7684\u566a\u58f0\u6c34\u5e73\u3002\u8bf7\u6ce8\u610f\uff0c\u8fd9\u91cc\u6211\u4eec\u53ea\u89c2\u5bdf\u56fe\u50cf\u5e8f\u5217 ot<\/sub>\u3001\u5956\u52b1 rt<\/sub> \u548c\u52a8\u4f5c at<\/sub>\uff0c\u6240\u6709\u5176\u4ed6\u90e8\u5206\u90fd\u662f\u6839\u636e\u6570\u636e\u4f30\u8ba1\u7684\uff0c\u5305\u62ec\u6f5c\u5728\u72b6\u6001 st<\/sub>\u3001\u72b6\u6001\u8f6c\u79fb\u51fd\u6570\u3001\u53d8\u5316\u56e0\u5b50 \u03b8 <\/strong>\u4ee5\u53ca\u53d8\u91cf\u4e4b\u95f4\u7684\u56fe\u7ed3\u6784\u3002<\/span><\/section>\n

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

<\/section>\n
\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u56fe\u8868\u793a\u63d0\u4f9b\u4e86\u4e00\u79cd\u7d27\u51d1\u7684\u65b9\u5f0f\u6765\u7f16\u7801\u53d8\u5316\u7684\u5185\u5bb9\u548c\u4f4d\u7f6e<\/span>\uff08what and where the changes ae\uff09<\/span>\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u9996\u5148\u4ece\u56fe\u7ed3\u6784\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u5f88\u76f4\u89c2\u5730\u770b\u51fa\u52a8\u4f5c\u9884\u6d4b\u9700\u8981\u54ea\u4e9b\u7ef4\u5ea6\u7684\u72b6\u6001\u3002\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u7b56\u7565\u4f18\u5316\u53ea\u9700\u8981 s1<\/sub> \u548c s3<\/sub>\uff0c\u800c\u4e0d\u9700\u8981 s2<\/sub>\uff0c\u56e0\u4e3a\u5b83\u4e0d\u4f1a\u5f71\u54cd\u672a\u6765\u7684\u5956\u52b1\u3002\u6b64\u5916\uff0c\u6211\u4eec\u4ece\u56fe\u8868\u793a\u91cc\u53ef\u4ee5\u5feb\u901f\u77e5\u9053\u53d8\u5316\u5728\u54ea\u91cc\uff0c\u5e76\u4e14\u77e5\u9053\u54ea\u4e9b\u53d8\u5316\u9700\u8981\u9002\u5e94\u3002\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u53ea\u6709 \u03b8s<\/sup> \u6355\u6349\u5230\u7684\u91cd\u529b\u53d8\u5316\u548c \u03b8r<\/sup> \u6355\u6349\u5230\u7684\u53ef\u80fd\u7684\u5956\u52b1\u53d8\u5316\u9700\u8981\u9002\u5e94\uff0c\u4f46\u6211\u4eec\u4e0d\u9700\u8981\u9002\u5e94 \u03b8o<\/sup> \u6355\u6349\u5230\u7684\u52a0\u5728\u56fe\u50cf\u4e0a\u7684\u566a\u58f0\u7684\u53d8\u5316\uff0c\u56e0\u4e3a\u5b83\u4e0d\u4f1a\u5f71\u54cd\u672a\u6765\u7684\u5956\u52b1\u3002\u53e6\u5916\uff0c\u7531\u4e8e\u5177\u6709\u6a21\u5757\u5316<\/span>\uff08modularity\uff09<\/span>\u7684\u7279\u6027\uff0c\u53ef\u4ee5\u5355\u72ec\u8003\u8651\u5404\u4e2a\u56e0\u7d20\uff0c\u907f\u514d\u4e86\u7ef4\u6570\u707e\u96be\u3002\u56e0\u6b64\uff0c\u901a\u8fc7\u660e\u786e\u5730\u5229\u7528\u8fd9\u79cd\u7d27\u51d1\u7684\u8868\u793a\u6765\u7f16\u7801\u53d8\u5316\uff0c\u5728Cartpole\u8fd9\u4e2a\u4f8b\u5b50\u4e2d\u5982\u4f55\u5e73\u8861pole\u7684\u7b56\u7565\u53ea\u968f\u7740\u91cd\u529b\u800c\u53d8\u5316\uff0c\u6240\u4ee5\u5728\u4e00\u4e2a\u65b0\u7684\u9886\u57df\uff0c\u53ea\u9700\u8981\u5c11\u91cf\u6837\u672c\u6765\u4f30\u8ba1\u4f4e\u7ef4\u53d8\u5316\u56e0\u5b50\uff0c\u7136\u540e\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u63a8\u5bfc\u51fa\u65b0\u57df\u4e2d\u7684\u6700\u4f18\u7b56\u7565\u3002\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u56fe\u77e5\u8bc6\u63d0\u4f9b\u4e86\u66f4\u53ef\u89e3\u91ca\u548c\u66f4\u6709\u6548\u7684\u9002\u5e94\u3002\u6b64\u5916\uff0c\u901a\u8fc7\u8bc6\u522b\u53d8\u5316\u7684\u4f4d\u7f6e\u548c\u65b9\u5f0f\uff0c\u6211\u4eec\u4e0d\u4ec5\u5229\u7528\u4e86\u4e0d\u53d8\u90e8\u5206\uff0c\u8fd8\u5229\u7528\u4e86\u53ef\u53d8\u90e8\u5206\uff0c\u56e0\u6b64\u53ef\u4ee5\u66f4\u6709\u6548\u5730\u9002\u5e94\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u4ec5\u4f9d\u9760\u4e0d\u53d8\u90e8\u5206\u53ef\u80fd\u4f1a\u727a\u7272\u90e8\u5206\u6027\u80fd\uff0c\u56e0\u4e3a\u4e00\u4e9b\u5173\u952e\u4fe1\u606f\u4f1a\u4e22\u5931\u3002<\/span><\/section>\n

<\/span><\/section>\n
References<\/span><\/strong><\/span>
<\/section>\n
B. Huang, F. Feng, C. Lu, S. Magliacane, and K. Zhang. Adarl: What, where, and how to adapt in transfer reinforcement learning. ICLR, 2022. <\/span><\/section>\n
Huang, Biwei, et al. “Causal Discovery from Heterogeneous\/Nonstationary Data.” J. Mach. Learn. Res. 21.89 (2020): 1-53.<\/span><\/section>\n
B. Huang, K. Zhang, M. Gong, and C. Glymour. Causal discovery and forecasting in nonstationary environments with state-space models. In ICML, 2019c. <\/span><\/section>\n
Zhang, Kun, et al. “Domain adaptation as a problem of inference on graphical models.” Advances in Neural Information Processing Systems 33 (2020): 4965-4976.<\/span><\/section>\n
Xie, Shaoan, et al. “Unaligned image-to-image translation by learning to reweight.” Proceedings of the IEEE\/CVF International Conference on Computer Vision. 2021.<\/span><\/section>\n

<\/section>\n


<\/p>\n

\n
\n
\n
\n
\n

<\/section>\n

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

4. \u76ee\u524d\u7684\u5206\u6790\u5de5\u5177\u4ee5\u53ca\u5bf9\u672a\u6765\u7684\u5c55\u671b<\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

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


<\/h2>\n

4.1 Causal-learn<\/span><\/strong><\/span><\/h2>\n

<\/section>\n
Causal-learn<\/strong>\uff0c\u7531CMU\u5f20\u5764\u8001\u5e08\u4e3b\u5bfc\uff0c\u591a\u4e2a\u56e2\u961f<\/span>\uff08CMU-CLeaR\u56e0\u679c\u7814\u7a76\u56e2\u961f\u3001\u8521\u745e\u521d\u8001\u5e08\u56e2\u961f\u3001\u5bab\u660e\u660e\u8001\u5e08\u56e2\u961f\u548cShohei Shimizu\u8001\u5e08\u56e2\u961f\uff09<\/span>\u8054\u5408\u5f00\u53d1\u51fa\u54c1\u7684\u56e0\u679c\u53d1\u73b0<\/strong>\u7b97\u6cd5\u5e73\u53f0\u3002<\/span><\/section>\n

<\/span><\/section>\n
Causal-learn\u7528Python\u5b9e\u73b0\u4e86CMU\u5f00\u53d1\u7684\u57fa\u4e8eJava\u7684Tetrad\u56e0\u679c\u53d1\u73b0\u5e73\u53f0\uff0c\u5e76\u8fdb\u4e00\u6b65\u52a0\u5165\u65b0\u7684\u7b97\u6cd5\u548c\u529f\u80fd\u3002\u5176\u4e2d\u5305\u542b\u4e86\u56e0\u679c\u53d1\u73b0\u7684\u7ecf\u5178\u7b97\u6cd5\u4e0eAPI\uff0c\u5e76\u4e14\u63d0\u4f9b\u4e86\u6a21\u5757\u5316\u7684\u4ee3\u7801\uff0c\u4ee5\u65b9\u4fbf\u7814\u7a76\u8005\u5b9e\u73b0\u81ea\u5df1\u7684\u7b97\u6cd5\u3002Causal-learn\u6240\u6709\u6a21\u5757\u5747\u57fa\u4e8ePython\u5b9e\u73b0\uff0c\u4ece\u800c\u907f\u514d\u4e86\u4f20\u7edf\u56e0\u679c\u53d1\u73b0\u5e93\u5bf9R\/Java\u7684\u4f9d\u8d56\uff0c\u4e3aPython\u5f00\u53d1\u8005\u63d0\u4f9b\u4fbf\u5229\u3002<\/span><\/section>\n

<\/span><\/section>\n
Causal-learn\u652f\u6301\uff1a<\/span><\/section>\n