{"id":813,"date":"2021-02-04T20:42:17","date_gmt":"2021-02-04T12:42:17","guid":{"rendered":"https:\/\/xg1990.com\/blog\/?p=813"},"modified":"2021-02-04T20:42:19","modified_gmt":"2021-02-04T12:42:19","slug":"%e5%81%8f%e8%a7%81%e4%b8%8e%e7%ae%97%e6%b3%95%e5%85%ac%e5%b9%b3-bias-and-algorithmic-fairness","status":"publish","type":"post","link":"https:\/\/xg1990.com\/blog\/archives\/813","title":{"rendered":"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 &#8211; Bias and Algorithmic Fairness"},"content":{"rendered":"\n<p>\u672c\u6587\u7ffb\u8bd1\u81ea\uff1ahttps:\/\/towardsdatascience.com\/bias-and-algorithmic-fairness-10f0805edc2b<\/p>\n\n\n\n<p><a href=\"https:\/\/medium.com\/m\/signin?actionUrl=%2F_%2Fbookmark%2Fp%2F10f0805edc2b&amp;operation=register&amp;redirect=https%3A%2F%2Ftowardsdatascience.com%2Fbias-and-algorithmic-fairness-10f0805edc2b&amp;source=post_actions_header--------------------------bookmark_preview-----------\"><\/a><\/p>\n\n\n\n<p>\u968f\u7740\u6570\u636e\u79d1\u5b66\u6cbf\u7740\u7092\u4f5c\u5468\u671f\u53d1\u5c55\uff0c\u5e76\u4f5c\u4e3a\u4e00\u79cd\u4e1a\u52a1\u529f\u80fd\u65e5\u8d8b\u6210\u719f\uff0c\u8be5\u5b66\u79d1\u9762\u4e34\u7684\u6311\u6218\u4e5f\u968f\u4e4b\u800c\u751f\u3002<\/p>\n\n\n\n<p>\u8fc7\u53bb\u51e0\u5e74\uff0c\u6570\u636e\u79d1\u5b66\u7684\u95ee\u9898\u9648\u8ff0\u4ece\u201c\u6211\u4eec\u6d6a\u8d3980%\u7684\u65f6\u95f4\u51c6\u5907\u6570\u636e\u201d\u5230\u201c\u751f\u4ea7\u90e8\u7f72\u662f\u6570\u636e\u79d1\u5b66\u6700\u56f0\u96be\u7684\u90e8\u5206\u201d\u518d\u5230\u201c\u7f3a\u4e4f\u53ef\u6d4b\u91cf\u7684\u4e1a\u52a1\u5f71\u54cd\u201d\u3002<\/p>\n\n\n\n<p>\u4f46\u662f\uff0c\u968f\u7740\u5546\u4e1a\u6570\u636e\u79d1\u5b66\u4f5c\u4e3a\u4e00\u79cd\u5546\u4e1a\u529f\u80fd\u7684\u6574\u5408\uff0c\u5e76\u514b\u670d\u4e86\u65e9\u671f\u7684\u95ee\u9898\uff0c\u6211\u4eec\u9762\u4e34\u7740\u65b0\u7684\u5177\u6709\u6311\u6218\u6027\u7684\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul><li>\u6570\u636e\u4f26\u7406\uff0c<\/li><li>\u6a21\u578b\u53ef\u89e3\u91ca\u6027\u548c\u95ee\u8d23\u5236\uff0c<\/li><li>\u7b97\u6cd5\u516c\u5e73\u6027\u3002<\/li><\/ul>\n\n\n\n<p>\u867d\u7136\u6570\u636e\u79d1\u5b66\u5bb6\u548c\u5546\u4e1a\u9886\u8896\u53ef\u4ee5\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4f9d\u8d56\u6280\u672f\u8fdb\u6b65\u6765\u89e3\u51b3\u7b2c\u4e00\u8f6e\u6570\u636e\u79d1\u5b66\u521d\u671f\u7684\u95ee\u9898\uff0c\u4f46\u4ec5\u4ec5\u6307\u671b\u6280\u672f\u6765\u89e3\u51b3\u8fd9\u4e9b\u65b0\u6311\u6218\u662f\u9519\u8bef\u7684\u3002<\/p>\n\n\n\n<p>\u6211\u4e2a\u4eba\u627f\u8ba4\uff0c\u6211\u5bf9\u5982\u4f55\u4f7f\u7528\u6570\u636e\u592a\u5929\u771f\u4e86\uff0c\u800c\u4e14\u5728\u5982\u4f55\u5904\u7406\u6570\u636e\u65b9\u9762\u53d7\u5230\u4e86\u76f8\u5f53\u5927\u7684\u8bef\u5bfc\u3002\u6211\u4e0d\u80fd\u4e00\u4e2a\u4eba\u627f\u8ba4\u5417\uff1f<\/p>\n\n\n\n<p>\u8fd9\u5c06\u4e0d\u662f\u4e00\u4e2a\u5b66\u672f\u4e0a\u548c\u6cd5\u5f8b\u4e0a\u5168\u9762\u7684\u8ba8\u8bba\u8fd9\u4e2a\u8bdd\u9898\uff0c\u800c\u662f\u6211\u4e2a\u4eba\u7684\u7ecf\u9a8c\u3002\u4e3a\u4e86\u66f4\u8be6\u7ec6\u5730\u4e86\u89e3\u4eba\u5de5\u667a\u80fd\u4e2d\u516c\u5e73\u548c\u8d23\u4efb\u7684\u6cd5\u5f8b\u542b\u4e49\uff0c\u6211\u5efa\u8bae\u9075\u5faaSandra Wachter\u535a\u58eb\u7684\u89c2\u70b9\uff0c\u9605\u8bfb\u201c\u5408\u7406\u63a8\u65ad\u7684\u6743\u5229\uff1a\u91cd\u65b0\u601d\u8003\u5927\u6570\u636e\u548c\u4eba\u5de5\u667a\u80fd\u65f6\u4ee3\u7684\u6570\u636e\u4fdd\u62a4\u6cd5\u201c<a href=\"https:\/\/www.law.ox.ac.uk\/business-law-blog\/blog\/2018\/10\/right-reasonable-inferences-re-thinking-data-protection-law-age-big\">A Right to Reasonable Inferences: Re-thinking Data Protection Law in the Age of Big Data and AI<\/a>\u201d<\/p>\n\n\n\n<h1 id=\"7428\">\u9ed1\u5323\u5b50\u6ca1\u4e86\uff1f<\/h1>\n\n\n\n<p>\u6211\u5bf9\u6570\u636e\u79d1\u5b66\u7684\u4f5c\u7528\u7684\u7406\u89e3\u9047\u5230\u4e86\u7b2c\u4e00\u4e2a\u6311\u6218\uff0c\u90a3\u5c31\u662f\u9010\u6e10\u6d88\u9000\u7684\u5ba3\u4f20\u4e0d\u518d\u80fd\u6210\u4e3a\u673a\u5668\u5b66\u4e60\u9ed1\u5323\u5b50\u7684\u501f\u53e3\u3002\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u4ece\u4e00\u4e2a\u610f\u60f3\u4e0d\u5230\u7684\u65b9\u5411\u8fdb\u5165\u4e86\u6211\u7684\u601d\u7ef4\uff1a\u6211\u7684\u5229\u76ca\u76f8\u5173\u8005\u3002\u5229\u76ca\u76f8\u5173\u8005\u6bd4\u5927\u591a\u6570\u6570\u636e\u79d1\u5b66\u5bb6\u66f4\u5173\u5fc3\u6a21\u578b\u7684\u8be6\u7ec6\u5de5\u4f5c\uff01\u8fd9\u5e76\u4e0d\u610f\u5473\u7740\u6211\u6ca1\u6709\u610f\u8bc6\u5230\u9ed1\u5323\u5b50\uff0c\u4f46\u6211\u4ece\u98ce\u9669\u7f13\u89e3\u7684\u5fc3\u6001\u6765\u5904\u7406\u5b83\u3002\uff08\u4ea4\u53c9\u9a8c\u8bc1\u3001\u5bf9\u6a21\u578b\u8f93\u5165\u548c\u8f93\u51fa\u7684\u5e7f\u6cdb\u6d4b\u8bd5\u3001\u65e5\u5fd7\u8bb0\u5f55\u548c\u76d1\u89c6\u3001\u5bf9\u6a21\u578b\u4f7f\u7528\u7684\u9650\u5236\u3001\u5927\u91cf\u7684\u5584\u610f\u3001\u5feb\u901f\u5931\u8d25\u5e76\u91cd\u8bd5\u2026\u2026\uff09<\/p>\n\n\n\n<p>\u6211\u505a\u8fc7\uff0c\u4f46\u4ecd\u7136\u628a\u6570\u636e\u79d1\u5b66\u89c6\u4e3a\u751f\u4ea7\u4e2d\u7684\u5b9e\u9a8c\u3002\u4f46\u662f\uff0c\u867d\u7136\u6211\u4ee5\u524d\u8ba4\u4e3a\u9ed1\u5323\u5b50\u53ea\u9700\u8981\u4fdd\u969c\u63aa\u65bd\u6765\u786e\u4fdd\u9884\u6d4b\u6709\u610f\u4e49\uff0c\u4f46\u6211\u73b0\u5728\u660e\u767d\u4e86\u4e3a\u4ec0\u4e48\u9700\u8981\u4e00\u4e9b\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\uff0c\u4e0d\u4ec5\u662f\u4e3a\u4e86\u95ee\u8d23\uff0c\u800c\u4e14\u662f\u4e3a\u4e86\u5bf9\u9884\u6d4b\u6709\u610f\u4e49\u6709\u5b9e\u9645\u7684\u4fe1\u5fc3\u3002\u4e3a\u4e86\u9650\u5236ML\u9ed1\u76d2\u7684\u4eba\u4e3a\u611a\u8822\uff0c\u6211\u4eec\u9700\u8981\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u6765\u5b9a\u4e49\u4f55\u65f6\u5bf9\u8868\u73b0\u826f\u597d\u7684\u8f93\u5165\u6570\u636e\u8fdb\u884c\u8bc4\u5206\u662f\u5b89\u5168\u7684\uff0c\u4f55\u65f6\u6700\u597d\u4e0d\u8fdb\u884c\u8bc4\u5206\uff0c\u5373\u4ee5\u524d\u770b\u4e0d\u5230\u7684\u8f93\u5165\u6570\u636e\u3002<\/p>\n\n\n\n<p><strong>ML\u975e\u5e38\u9002\u5408\u548c\u5f3a\u5927\u7684\u51b3\u7b56\u7a7a\u95f4\u63d2\u503c\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u4f7f\u7528ML\u5c06\u51b3\u7b56\u63a8\u65ad\u5230\u65b0\u7684\u548c\u770b\u4e0d\u89c1\u7684\u6570\u636e\u7684\u4e0d\u5b89\u5168\u533a\u57df\u592a\u5bb9\u6613\u548c\u8bf1\u4eba\u4e86\uff0c\u8fd9\u4ece\u6765\u4e0d\u662f\u6a21\u578b\u8bad\u7ec3\u548c\u6a21\u578b\u9a8c\u8bc1\u7684\u4e00\u90e8\u5206\u3002<\/strong><\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu\" alt=\"Image for post\"\/><figcaption><a href=\"https:\/\/xkcd.com\/605\/\"><strong>https:\/\/xkcd.com\/605\/<\/strong><\/a><strong>, (CC BY-NC 2.5)<\/strong><\/figcaption><\/figure><\/div>\n\n\n\n<p id=\"2795\">\u867d\u7136ML\u9ed1\u5323\u5b50\u5728\u6211\u7684\u65e5\u5e38\u5de5\u4f5c\u4e2d\u53ea\u662f\u4e00\u4e2a\u7070\u8272\u7684\u76d2\u5b50\uff0c\u4f46\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u4e3a\u6211\u63d0\u4f9b\u4e86\u5fc5\u8981\u7684\u4fe1\u5fc3\uff0c\u5e76\u63d0\u9ad8\u4e86\u6211\u4eec<strong>\u6291\u5236\u4eba\u4e3a\u611a\u8822<\/strong>\u7684\u80fd\u529b\uff0c\u4f7f\u6211\u4eec\u7684\u6a21\u578b\u4fdd\u6301\u5728\u4f17\u6240\u5468\u77e5\u7684\u5b89\u5168\u4fdd\u969c\u8303\u56f4\u5185\u3002\u8fd9\u610f\u5473\u7740\u6709\u65f6\u6700\u597d\u4e0d\u8981\u63d0\u4f9b\u5206\u6570\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter\"><img src=\"https:\/\/miro.medium.com\/max\/640\/0*7ETHreMJSHm1avre.png\" alt=\"Image for post\"\/><figcaption><a href=\"https:\/\/www.google.com\/url?sa=i&amp;source=images&amp;cd=&amp;cad=rja&amp;uact=8&amp;ved=2ahUKEwiW45rBooDlAhVFJBoKHdhsDNMQjhx6BAgBEAI&amp;url=http%3A%2F%2Fscikit-learn.org%2Fstable%2Fmodules%2Fmixture.html&amp;psig=AOvVaw38TRN5Wp0pRmqLDX7USmfB&amp;ust=1570198236469560\">Scikit<\/a><\/figcaption><\/figure><\/div>\n\n\n\n<p>\u5b9e\u73b0\u8fd9\u4e00\u70b9\u7684\u4e00\u4e2a\u975e\u5e38\u7b80\u5355\u4f46\u975e\u5e38\u6709\u6548\u7684\u65b9\u6cd5\u662f\uff0c\u5728\u8bad\u7ec3\u6570\u636e\u4e0a\u62df\u5408\u4e00\u4e2a\u72ec\u7acb\u7684<strong>\u9ad8\u65af\u6df7\u5408\u6a21\u578b<\/strong>\uff0c\u4ee5\u786e\u5b9a\u51b3\u7b56\u7a7a\u95f4\u7684\u7f6e\u4fe1\u8fb9\u754c\u3002\u5982\u679c\u60a8\u7684\u6a21\u578b\u8f93\u5165\u4e0d\u5728GMM\u7684\u7f6e\u4fe1\u8303\u56f4\u5185\uff0c\u5c31\u4e0d\u8981\u8fd4\u56de\u6a21\u578b\u8f93\u51fa\u3002<\/p>\n\n\n\n<p>\u76ee\u524d\uff0c\u6211\u5e0c\u671b\uff0c\u4e00\u65b9\u9762\uff0c\u6570\u636e\u79d1\u5b66\u9886\u57df\u7684\u79d1\u6280\u8fdb\u6b65\u5c06\u8fdb\u4e00\u6b65\u89e3\u51b3\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u95ee\u9898\uff0c\u53e6\u4e00\u65b9\u9762\uff0c\u4f5c\u4e3a\u4e00\u540d\u6570\u636e\u79d1\u5b66\u5bb6\uff0c\u6211\u5f88\u4e50\u610f\u627f\u62c5\u6a21\u578b\u53ef\u89e3\u91ca\u6027\u7684\u8d23\u4efb\u3002<\/p>\n\n\n\n<h1 id=\"0d02\">\u504f\u89c1<\/h1>\n\n\n\n<p>\u867d\u7136\u6a21\u578b\u7684\u53ef\u89e3\u91ca\u6027\u662f\u5546\u4e1a\u6570\u636e\u79d1\u5b66\u804c\u80fd\u90e8\u95e8\u672c\u8eab\u7684\u8d23\u4efb\uff0c\u4f46\u6570\u636e\u4f26\u7406\u548c\u7b97\u6cd5\u516c\u5e73\u6027\u4e0d\u80fd\u4ec5\u7531\u6570\u636e\u79d1\u5b66\u90e8\u95e8\u62e5\u6709\uff01\u4e3a\u4ec0\u4e48\uff1f\u56e0\u4e3a\u516c\u5e73\u6a21\u578b\u662f\u5bf9\u7cbe\u786e\u6a21\u578b\u7684\u4e00\u79cd\u6743\u8861\uff0c\u4fdd\u6301\u8fd9\u79cd\u5e73\u8861\u610f\u5473\u7740\u5b83\u5c06\u6210\u4e3a\u9ad8\u7ea7\u9886\u5bfc\u7684\u8d23\u4efb\u3002\u5e0c\u671b\u5f53\u4f60\u5728\u6587\u7ae0\u7684\u7ed3\u5c3e\u770b\u5230\u603b\u7ed3\u7684\u65f6\u5019\uff0c\u8fd9\u4f1a\u53d8\u5f97\u66f4\u6709\u610f\u4e49\u3002<\/p>\n\n\n\n<p>\u73b0\u5b9e\u60c5\u51b5\u662f\uff0c\u4efb\u4f55\u7531\u4eba\u7c7b\u884c\u4e3a\u4ea7\u751f\u6216\u884d\u751f\u7684\u6570\u636e\u70b9\uff0c\u90fd\u4f1a\u5728\u672c\u8d28\u4e0a\u53d7\u5230\u6211\u4eec\u4eba\u7c7b\u504f\u89c1\u4f3c\u4e4e\u65e0\u7a77\u65e0\u5c3d\u7684\u5217\u8868\u7684\u6c61\u67d3\u3002\u53ea\u662f\u60f3\u8bf4\u51e0\u53e5\uff1a<\/p>\n\n\n\n<p><strong>\u7fa4\u4f53\u5185\u7684\u504f\u8892\u548c\u7fa4\u4f53\u5916\u7684\u6d88\u6781<\/strong>\uff1a\u504f\u597d\u6211\u4eec\u81ea\u5df1\u7684\u793e\u4f1a\u7fa4\u4f53\u4e2d\u7684\u4eba\uff0c\u800c\u4e0d\u662f\u503e\u5411\u4e8e\u60e9\u7f5a\u6216\u7ed9\u7fa4\u4f53\u5916\u7684\u4eba\u65bd\u52a0\u8d1f\u62c5\u3002\u504f\u89c1\u5728\u504f\u89c1\u548c\u6b67\u89c6\u4e2d\u8d77\u7740\u91cd\u8981\u4f5c\u7528\u3002<\/p>\n\n\n\n<p><strong>\u57fa\u672c\u5f52\u56e0\u9519\u8bef<\/strong>\uff1a\u5f53\u6211\u4eec\u503e\u5411\u4e8e\u5c06\u67d0\u4eba\u7684\u884c\u4e3a\u5f52\u56e0\u4e8e\u5176\u6027\u683c\u7684\u5185\u5728\u54c1\u8d28\u800c\u4e0d\u662f\u60c5\u5883\u80cc\u666f\u65f6\u3002<\/p>\n\n\n\n<p><strong>\u6d88\u6781\u504f\u89c1<\/strong>\uff1a\u5f53\u6211\u4eec\u5f3a\u8c03\u6d88\u6781\u7ecf\u9a8c\u591a\u4e8e\u79ef\u6781\u7ecf\u9a8c\u65f6\uff0c\u8fd9\u5bf9\u793e\u4f1a\u5224\u65ad\u548c\u5370\u8c61\u5f62\u6210\u6709\u91cd\u8981\u5f71\u54cd\u3002<\/p>\n\n\n\n<p><strong>\u523b\u677f\u5370\u8c61<\/strong>\uff1a\u5f53\u6211\u4eec\u671f\u671b\u4e00\u4e2a\u7fa4\u4f53\u4e2d\u7684\u67d0\u4e2a\u6210\u5458\u5177\u6709\u67d0\u4e9b\u7279\u5f81\uff0c\u800c\u6ca1\u6709\u5173\u4e8e\u8fd9\u4e2a\u4eba\u7684\u5b9e\u9645\u4fe1\u606f\u65f6\u3002<\/p>\n\n\n\n<p><strong>\u968f\u6ce2\u9010\u6d41\u6548\u5e94<\/strong>\uff1a\u5f53\u6211\u4eec\u505a\u67d0\u4e8b\u6216\u76f8\u4fe1\u67d0\u4e8b\u65f6\uff0c\u56e0\u4e3a\u8bb8\u591a\u4eba\u90fd\u8fd9\u6837\u505a\u3002<\/p>\n\n\n\n<p><strong>\u504f\u89c1\u76f2\u70b9<\/strong>\uff1a\u6211\u4eec\u503e\u5411\u4e8e\u770b\u4e0d\u5230\u81ea\u5df1\u7684\u4e2a\u4eba\u504f\u89c1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/1600\/0*R1AFqFSih42NRTMl\" alt=\"Image for post\"\/><\/figure>\n\n\n\n<p id=\"c5b3\">\u6211\u4eec\u6240\u6709\u7684\u6570\u636e\u90fd\u662f\u6709\u504f\u89c1\u7684\uff0c\u56e0\u4e3a\u4e16\u754c\u662f\u6709\u504f\u89c1\u7684\u3002<\/p>\n\n\n\n<p><strong>\u4fee\u6b63\u4e16\u754c\u4e0a\u7684\u504f\u89c1\u8d85\u51fa\u4e86\u6570\u636e\u79d1\u5b66\u5bb6\u7684\u5de5\u4f5c\u89c4\u8303\u3002\u7136\u800c\uff0c\u4e0d\u4f7f\u7528ML\u6a21\u578b\u6765\u5f3a\u5316\u73b0\u6709\u7684\u504f\u89c1\u662f\u4e00\u4e2a\u6570\u5b66\u95ee\u9898\uff0c\u5b83\u5f88\u5feb\u5c31\u4ece\u5355\u7eaf\u7684\u9053\u5fb7\u4e49\u52a1\u8f6c\u53d8\u4e3a\u73b0\u4ee3\u6570\u636e\u79d1\u5b66\u5bb6\u7684\u5de5\u4f5c\u8981\u6c42\u3002<\/strong><\/p>\n\n\n\n<p id=\"4d15\">\u89e3\u51b3\u6570\u636e\u79d1\u5b66\u4e2d\u7684\u504f\u89c1\u662f\u4e00\u4e2a\u6781\u5176\u590d\u6742\u7684\u8bdd\u9898\uff0c\u6700\u91cd\u8981\u7684\u662f\u6ca1\u6709\u901a\u7528\u7684\u89e3\u51b3\u65b9\u6848\u6216\u94f6\u5f39\u3002\u5728\u4efb\u4f55\u4e00\u4f4d\u6570\u636e\u79d1\u5b66\u5bb6\u81f4\u529b\u4e8e\u51cf\u5c11\u504f\u89c1\u4e4b\u524d\uff0c\u6211\u4eec\u9700\u8981\u901a\u8fc7\u67e5\u9605\u4ee5\u4e0b<strong>\u516c\u5e73\u6027\u6811<\/strong>\uff0c\u5728\u6211\u4eec\u7684\u4e1a\u52a1\u95ee\u9898\u80cc\u666f\u4e0b\u5b9a\u4e49\u516c\u5e73\u6027\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/1218\/0*WMn7QPh8Nrds3Q-Z\" alt=\"Image for post\"\/><figcaption>The Fairness Tree, Adapted from <a href=\"http:\/\/www.datasciencepublicpolicy.org\/projects\/aequitas\/\">http:\/\/www.datasciencepublicpolicy.org\/projects\/aequitas\/<\/a><\/figcaption><\/figure>\n\n\n\n<p>\u6839\u636e\u60a8\u7684\u4e1a\u52a1\u95ee\u9898\u9605\u8bfb\u516c\u5e73\u6811\u5e76\u975e\u6613\u4e8b\u3002\u4f5c\u4e3a\u4e00\u4e2a\u4f8b\u5b50\uff08\u6539\u7f16\u81ea<a href=\"https:\/\/pair-code.github.io\/what-if-tool\/ai-fairness.html\">\u8fd9\u91cc<\/a>\uff09\uff0c\u5047\u8bbe\u4f60\u60f3\u8bbe\u8ba1\u4e00\u4e9bML\u7cfb\u7edf\u6765\u5904\u7406<strong>\u62b5\u62bc\u8d37\u6b3e\u7533\u8bf7<\/strong>\uff0c\u53ea\u6709\u4e00\u5c0f\u90e8\u5206\u7684\u7533\u8bf7\u662f\u7531\u5973\u6027\u63d0\u51fa\u7684\u3002<\/p>\n\n\n\n<ol><li><strong>\u7fa4\u4f53\u65e0\u610f\u8bc6\u9009\u62e9<\/strong>\uff1a\u6211\u4eec\u5728\u7533\u8bf7\u8fc7\u7a0b\u4e2d\u5b8c\u5168\u5ffd\u7565\u4e86\u6027\u522b\u4fe1\u606f\u3002\u56e0\u4e3a\u8d37\u6b3e\u4eba\u53ef\u4ee5\u6388\u4e88\u7684\u6279\u51c6\u6570\u91cf\u6709\u9650\uff0c\u6240\u4ee5\u6839\u636e\u5ba2\u89c2\u7684\u3001\u4e0d\u5206\u6027\u522b\u7684\u6807\u51c6\uff0c\u8fd9\u4e9b\u6279\u51c6\u53ea\u80fd\u6388\u4e88\u6700\u5408\u683c\u7684\u4e2a\u4eba\u3002\u4f46\u662f\uff0c\u5220\u9664\u6027\u522b\u548c\u6027\u522b\u4ee3\u7406\u4fe1\u606f\u5e76\u4e0d\u80fd\u89e3\u51b3\u5386\u53f2\u504f\u89c1\uff0c\u800c\u4e14\u901a\u5e38\u4e0d\u662f\u4e00\u4e2a\u975e\u5e38\u6709\u6548\u7684\u8fc7\u7a0b\u6765\u51cf\u8f7b\u504f\u89c1\uff0c\u6211\u4eec\u5c06\u5728\u4e0b\u9762\u7684\u7f8e\u56fd\u4eba\u53e3\u666e\u67e5\u6570\u636e\u7684\u5de5\u4f5c\u793a\u4f8b\u4e2d\u770b\u5230\u8fd9\u4e00\u70b9\u3002<\/li><li><strong>\u8c03\u6574\u540e\u7684\u7fa4\u4f53\u9608\u503c<\/strong>\uff1a\u7531\u4e8e\u5386\u53f2\u504f\u89c1\u4f7f\u5973\u6027\u770b\u8d77\u6765\u4e0d\u5982\u7537\u6027\u503c\u5f97\u8d37\u6b3e\uff0c\u4f8b\u5982\u5de5\u4f5c\u7ecf\u5386\u548c\u6258\u513f\u8d23\u4efb\uff0c\u6211\u4eec\u6309\u7fa4\u4f53\u4f7f\u7528\u4e0d\u540c\u7684\u6279\u51c6\u9608\u503c\u3002<\/li><li><strong>\u4eba\u53e3\u5747\u7b49<\/strong>\uff1a\u6279\u51c6\u7387\u5e94\u53cd\u6620\u5404\u7ec4\u7533\u8bf7\u7684\u767e\u5206\u6bd4\u3002\u4f46\u8fd9\u5e76\u6ca1\u6709\u8003\u8651\u5230\u62b5\u62bc\u8d37\u6b3e\u8fdd\u7ea6\u7684\u98ce\u9669\u3002<\/li><li><strong>\u5e73\u7b49\u673a\u4f1a<\/strong>\uff1a\u540c\u6837\u6bd4\u4f8b\u7684\u6709\u8d37\u6b3e\u4ef7\u503c\u7684\u7537\u5973\u83b7\u5f97\u4e86\u62b5\u62bc\u8d37\u6b3e\u3002\u8fd9\u4f3c\u4e4e\u7b26\u5408\u62b5\u62bc\u8d37\u6b3e\u673a\u6784\u7684\u4e1a\u52a1\u76ee\u6807\uff0c\u800c\u4e14\u4f3c\u4e4e\u662f\u516c\u5e73\u7684\u3002\u201c\u6709\u8d44\u683c\u83b7\u5f97\u7406\u60f3\u7ed3\u679c\u7684\u4e2a\u4eba\u5e94\u8be5\u6709\u5e73\u7b49\u7684\u673a\u4f1a\u88ab\u6b63\u786e\u5206\u7c7b\u4e3a\u8fd9\u79cd\u7ed3\u679c\u3002\u201d\uff08\u83ab\u91cc\u8328\u00b7\u54c8\u7279\uff09<\/li><li><strong>\u7cbe\u786e\u5e73<\/strong>\u7b49\uff1a\u4e0d\u53d1\u653e\u8d37\u6b3e\u4f1a\u5bf9\u4e2a\u4eba\u4ea7\u751f\u975e\u5e38\u8d1f\u9762\u7684\u5f71\u54cd\u3002\u5728\u673a\u4f1a\u5747\u7b49\u7684\u60c5\u51b5\u4e0b\uff0c\u4e24\u4e2a\u7fa4\u4f53\u90fd\u6709\u771f\u6b63\u7684\u6b63\u5747\u7b49\u3002\u4f46\u662f\u5982\u679c\u6a21\u578b\u9519\u8bef\u7684\u662f\u5973\u6027\u4e0d\u507f\u8fd8\u8d37\u6b3e\u7684\u9891\u7387\u662f\u7537\u6027\u7684\u4e24\u500d\uff08\u5047\u9634\u6027\uff09\uff0c\u90a3\u4e48\u6a21\u578b\u4f1a\u62d2\u7edd\u63a5\u53d7\u8d37\u6b3e\u7684\u5973\u6027\u662f\u7537\u6027\u7684\u4e24\u500d\u3002\u56e0\u6b64\uff0c\u5e94\u8be5\u8c03\u6574\u6a21\u578b\uff0c\u4f7f\u6a21\u578b\u9519\u8bef\u7684\u6b21\u6570\u5728\u4e24\u7ec4\u7684\u6279\u51c6\u548c\u62d2\u7edd\u603b\u6570\u4e2d\u6240\u5360\u7684\u767e\u5206\u6bd4\u76f8\u540c\u3002\u7cbe\u5ea6\u5947\u5076\u6821\u9a8c\u4e5f\u662f\u516c\u5e73\u6811\u7684\u5efa\u8bae\uff08\u9519\u8bef>\u60e9\u7f5a>\u5c0f\u5e72\u9884\u91cf\uff09\u3002<\/li><\/ol>\n\n\n\n<h1 id=\"adc3\">\u7b97\u6cd5\u516c\u5e73<\/h1>\n\n\n\n<p>\u786e\u5b9a\u516c\u5e73\u7684\u9002\u5f53\u5b9a\u4e49\u53ea\u662f\u7b2c\u4e00\u6b65\u3002\u4e0b\u4e00\u6b65\u662f\u4e3a\u504f\u5dee\u9009\u62e9\u5408\u9002\u7684\u7f13\u89e3\u7b56\u7565\u3002\u540c\u6837\uff0c\u51cf\u8f7b\u504f\u89c1\u662f\u4e00\u4e2a\u590d\u6742\u7684\u8bdd\u9898\uff0c\u540c\u6837\uff0c\u6ca1\u6709\u901a\u7528\u7684\u65b9\u6cd5\u6216\u94f6\u5f39\u3002\u4e0b\u56fe\u663e\u793a\u4e86IBM\u7684<a href=\"https:\/\/github.com\/IBM\/AIF360\" target=\"_blank\" rel=\"noreferrer noopener\">AIF360<\/a>\u5de5\u5177\u5305\u4e2d\u7f13\u89e3\u7b56\u7565\u548c\u73b0\u6709\u5b9e\u73b0\u76843\u4e2a\u4e3b\u8981\u7c7b\u522b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/1121\/0*qdtylavNHR7jWy5u\" alt=\"Image for post\"\/><\/figure>\n\n\n\n<p id=\"beb5\">IBM\u7684trusted AI\u662f\u4e86\u89e3\u66f4\u591a\u7ec6\u8282\u7684\u7edd\u4f73\u8d44\u6e90\uff1ahttps:\/\/www.research.ibm.com\/artificial-intelligence\/trusted-ai\/<\/p>\n\n\n\n<h2 id=\"57e4\">\u4f7f\u7528TensorFlow 2.0\u8fdb\u884c\u5bf9\u6297\u6027\u53bb\u504f\u89c1\u2014\u2014\u4e00\u4e2a\u4f8b\u5b50<\/h2>\n\n\n\n<p>\u4e3a\u4e86\u7a81\u51fa\u504f\u9887\u6570\u636e\u7684\u6311\u6218\u4ee5\u53ca\u6570\u636e\u79d1\u5b66\u5bb6\u80fd\u505a\u4e9b\u4ec0\u4e48\uff0c\u6211\u4eec\u6765\u770b\u770b1994\u5e74\u7f8e\u56fd\u4eba\u53e3\u666e\u67e5\u6536\u5165\u6570\u636e\u96c6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/1600\/0*0vQ2w8lDEif6BbDm\" alt=\"Image for post\"\/><\/figure>\n\n\n\n<p id=\"a9f4\">\u76ee\u6807\u662f\u5efa\u7acb\u4e00\u4e2a\u5206\u7c7b\u6a21\u578b\u6765\u9884\u6d4b\u4e00\u4e2a\u4eba\u7684\u5e74\u6536\u5165\u662f\u5426\u8d85\u8fc75\u4e07\u7f8e\u5143\u3002\u8fd9\u4e2a\u95ee\u9898\u4ece\u6839\u672c\u4e0a\u8bb2\u662f\u6709\u504f\u89c1\u7684\uff08\u6027\u522b\u5de5\u8d44\u5dee\u8ddd\u3001\u6b67\u89c6\u7b49\uff09\uff0c\u6570\u636e\u96c6\u5305\u62ec\u654f\u611f\u6570\u636e\uff1a<strong>\u79cd\u65cf\u548c\u6027\u522b<\/strong>\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4ece\u6a21\u578b\u8f93\u5165\u4e2d\u5220\u9664\u8fd9\u4e9b\u654f\u611f\u6570\u636e\uff0c\u8bd5\u56fe\u6784\u5efa\u4e00\u4e2a<strong>\u4e0d\u5206\u7ec4\u7684<\/strong>\u5206\u7c7b\u5668\u3002\u4e0b\u56fe\u663e\u793a\u4e86\u6211\u4eec\u7528\u4e8e\u89e3\u51b3\u95ee\u9898\u7684\u795e\u7ecf\u7f51\u7edc\u4f53\u7cfb\u7ed3\u6784\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/724\/0*TkUeXGNrjxB1R8pB\" alt=\"Image for post\"\/><\/figure>\n\n\n\n<p id=\"cdf8\">\u8be5\u4f53\u7cfb\u7ed3\u6784\u4f7f\u7528\u4e00\u4e2a\u7f16\u7801\u5668\u795e\u7ecf\u7f51\u7edc\u6765\u521b\u5efa\u4e00\u4e2a\u5171\u4eab\u7684\u6570\u636e\u5d4c\u5165\uff0c\u5b83\u4e3a3\u4e2a\u5206\u7c7b\u795e\u7ecf\u7f51\u7edc\u63d0\u4f9b\u6570\u636e\uff0c\u8fd9\u4e9b\u795e\u7ecf\u7f51\u7edc\u7684\u76ee\u6807\u662f\u9884\u6d4b\u4e2a\u4eba\u7684\u5de5\u8d44\u3001\u6027\u522b\u548c\u79cd\u65cf\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<pre id=\"cdf8\" class=\"wp-block-code\"><code lang=\"python\" class=\"language-python\">class AdversarialModel(tf.keras.Model):\n     def <strong>init<\/strong>(\n         self,\n         tf_feature_columns,\n         encoder_units, decoder_units, model_head_names\n     ):\n         super(AdversarialModel, self).<strong>init<\/strong>()\n         self.output_model_names = model_head_names\n         self.encoder_model = tf.keras.Sequential(\n             [\n                 tf.keras.layers.DenseFeatures(tf_feature_columns),\n                 tf.keras.layers.Dense(units=encoder_units, activation=tf.nn.leaky_relu),\n                 tf.keras.layers.Dropout(0.5),\n                 tf.keras.layers.Dense(units=encoder_units, activation=tf.nn.leaky_relu)\n             ], name=\"encoder\" \n         )\n <code>    for model_name in self.output_model_names:<\/code>\n<code>         model = tf.keras.Sequential(<\/code>\n<code>             [<\/code>\n<code>                 tf.keras.layers.InputLayer(input_shape=(encoder_units,)),<\/code>\n<code>                 tf.keras.layers.Dense(units=decoder_units,<\/code>\n<code> activation=tf.nn.leaky_relu),<\/code>\n<code>                 tf.keras.layers.Dropout(0.5),<\/code>\n<code>                 tf.keras.layers.Dense(units=decoder_units, activation=tf.nn.leaky_relu),<\/code>\n<code>                 tf.keras.layers.Dense(units=1, activation='linear', name=model_name),<\/code>\n<code>             ], name=model_name<\/code>\n<code>          )<\/code>\n<code>         setattr(self, model_name, model)<\/code>\n<code> def call(self, x):<\/code>\n<code>     x_enc = self.encoder_model(x)<\/code>\n<code>     return list(<\/code>\n<code>         map(<\/code>\n<code>             lambda model_name: getattr(self, model_name)(x_enc), self.output_model_names<\/code>\n<code>         )<\/code>\n<code>     )<\/code>\n# Create categorical embedding feature columns\n CATEGORICAL_COLUMNS = ['workclass', 'education', 'marital-status', 'occupation','relationship', 'native-country']\n NUMERIC_COLUMNS = ['age', 'education-num', 'capital-gain', 'capital-loss', 'hours-per-week']        \n tf_feature_columns = []\n for feature_name in CATEGORICAL_COLUMNS:\n   vocab = df[feature_name].unique()\n   tf_feature_columns.append(\n       tf.feature_column.embedding_column(\n           categorical_column=tf.feature_column.categorical_column_with_vocabulary_list(feature_name, vocab),\n           dimension=int(max([2, len(vocab)**0.25]))\n       )\n   )\n for feature_name in NUMERIC_COLUMNS:\n   tf_feature_columns.append(tf.feature_column.numeric_column(feature_name, dtype=tf.float32))\n model = AdversarialModel(tf_feature_columns, 128, 64, [\"salary\", \"sex\", \"race\"])\n model.summary()\n \"\"\"\n Model: \"adversarial_model\"\n \n Layer (type)                 Output Shape              Param #   \n encoder (Sequential)         multiple                  19006     \n \n salary (Sequential)          (None, 1)                 12481     \n \n sex (Sequential)             (None, 1)                 12481     \n \n race (Sequential)            (None, 1)                 12481     \n Total params: 56,449\n Trainable params: 56,449\n Non-trainable params: 0\n \"\"\"<\/code><\/pre>\n\n\n\n<p id=\"61ec\">\u5728\u7b2c\u4e00\u8f6e\u4e2d\uff0c\u6211\u4eec<strong>\u4e0d\u4f7f\u7528<\/strong>\u4efb\u4f55\u5bf9\u6297\u6027\u68af\u5ea6\uff0c\u800c\u662f\u5c06\u6a21\u578b\u8bad\u7ec3\u4e3a\u6807\u51c6\u7684<strong>\u591a\u7c7b\u5206\u7c7b\u5668<\/strong>\uff0c\u4ee5\u89c2\u5bdf\u6211\u4eec\u5c1d\u8bd5\u7684\u7fa4\u4f53\u65e0\u610f\u8bc6\u662f\u591a\u4e48\u6210\u529f\u3002\u4e0b\u8868\u663e\u793a\u4e8620\u8f6e\u8bad\u7ec3\u540e\u7684\u6a21\u578b\u6027\u80fd\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/1600\/0*zZ3ZVcC3qVEbD2Ma\" alt=\"Image for post\"\/><figcaption>Epoch 20: Loss: 1.357 salary_accuracy: 0.829 sex_accuracy: 0.825 race_accuracy: 0.830<\/figcaption><\/figure>\n\n\n\n<p>\u663e\u7136\uff0c<strong>\u5220\u9664\u654f\u611f\u6570\u636e\u5e76\u4e0d\u662f\u8ba9\u6211\u4eec\u7684\u6a21\u578b\u62b9\u53bb\u504f\u89c1<\/strong>\u3002\u4ece\u5269\u4e0b\u7684\u6570\u636e\u4e2d\u6211\u4eec\u53ef\u4ee5\u5f88\u597d\u5730\u9884\u6d4b\u6027\u522b\u548c\u79cd\u65cf\u3002\u8fd9\u4e00\u70b9\u4e5f\u4e0d\u5947\u602a\uff0c\u56e0\u4e3a\u6211\u4eec\u7684\u95ee\u9898\u4e5f\u6709\u5386\u53f2\u504f\u89c1\uff0c\u654f\u611f\u6570\u636e\u4e0e\u6240\u6709\u5176\u4ed6\u7279\u5f81\u76f8\u5173\uff0c\u4f8b\u5982\u6559\u80b2\u3001\u804c\u4e1a\u9009\u62e9\u7b49\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5c06\u4f7f\u7528<strong>\u5bf9\u6297\u6027\u795e\u7ecf\u7f51\u7edc<\/strong>\u4ece\u5171\u4eab\u6570\u636e\u5d4c\u5165\u4e2d\u5220\u9664\u4efb\u4f55\u79cd\u65cf\u6216\u6027\u522b\u76f8\u5173\u4fe1\u606f\u3002TensorFlow 2.0\u7684\u81ea\u5b9a\u4e49\u8bad\u7ec3\u5faa\u73af\u5bf9NN\u5934\u8fdb\u884c4\u4e2a\u65f6\u671f\u7684\u8bad\u7ec3\uff0c\u4ee5\u9002\u5e94\u5f53\u524d\u7684\u5171\u4eab\u6570\u636e\u5d4c\u5165\uff0c\u7136\u540e\u4f7f\u7528\u6027\u522b\u548c\u79cd\u65cf\u5206\u7c7b\u5668\u7684\u8d1f\u68af\u5ea6\u5bf9\u7f16\u7801\u5668\u8fdb\u884c\u5bf9\u6297\u6027\u8bad\u7ec3\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code lang=\"python\" class=\"language-python\">def train_one_batch(model, features, targets, label_imbalance, loss_weights, model_name, encoder_loss_signs):\n     with tf.GradientTape() as tape:\n         logits = model(features)\n         y_ = tf.math.sigmoid(logits)\n         nn_head_losses = list(\n             map(\n                 lambda i: tf.nn.weighted_cross_entropy_with_logits(\n                     labels=tf.expand_dims(targets[i], 1),\n                     logits=logits[i],\n                     pos_weight=np.array([label_imbalance[i], 1.0\/label_imbalance[i]])\n                 ), range(len(model.output_model_names))\n             )\n         )\n         encoder_loss = tf.add_n(\n             list(\n                 map(\n                     lambda x: x[0]<em>x[1]<\/em>x[2], list(zip(encoder_loss_signs, loss_weights, nn_head_losses))\n                 )\n             )\n         )\n         if model_name == \"encoder\":\n             grad = tape.gradient(encoder_loss, model.trainable_variables)\n         else:\n             model_index = model.output_model_names.index(model_name)\n             grad = tape.gradient(nn_head_losses[model_index], model.trainable_variables)\n     return y_, encoder_loss, grad\n Assemble a TF dataset\n sensitives_sex = (df_train['sex'] == 'Female').astype(dtype=np.float32)\n sensitives_race = (df_train['race'] != 'White').astype(np.float32)\n label = (df_train['label'] == '>50K').astype(dtype=np.float32)\n BATCH_SIZE = 100\n ds_train = tf.data.Dataset.from_tensor_slices(\n     (\n         (df_train.loc[:,CATEGORICAL_COLUMNS+NUMERIC_COLUMNS]).to_dict(\"list\"),\n         (label.values, sensitives_sex.values, sensitives_race.values)\n     )\n ).batch(BATCH_SIZE)\n epochs = 100\n labels = ['salary','sex', 'race']\n loss_weights = [0.5, 0.25, 0.25] # The weights of the NN heads in the loss of the Encoder\n encoder_loss_signs = [1.0, -1.0, -1.0] # Negative adversarial gradients for sex and race\n optimizer = tf.keras.optimizers.Adam(learning_rate=1e-5)\n Our manual training loop\n for epoch in range(epochs):\n     accuracy = dict(\n         map(\n             lambda x: (x, tf.keras.metrics.BinaryAccuracy(threshold=0.5)), labels\n         )\n     )\n     epoch_loss_avg = tf.keras.metrics.Mean()\n     is_adversarial_emb_training = epoch % 5 == 0\n     for features, targets in ds_train:\n         label_imbalance = list(map(lambda x: (x.numpy() == 1).sum() \/ len(x.numpy()), targets))\n         if is_adversarial_emb_training:\n             model.encoder_model.trainable = True\n             for output_model_name in model.output_model_names:\n                 getattr(model, output_model_name).trainable = False\n             <em>, loss, grad = train_one_batch(<\/em>\n<em>                 model, features, targets, label_imbalance,<\/em>\n<em>                 loss_weights, \"encoder\", encoder_loss_signs)<\/em>\n<em>             optimizer.apply_gradients(zip(grad, model.trainable_variables))<\/em>\n<em>         model.encoder_model.trainable = False<\/em>\n<em>         for train_output_model_index, train_output_model_name in enumerate(model.output_model_names):<\/em>\n<em>             for output_model_name in model.output_model_names:<\/em>\n<em>                 if output_model_name == train_output_model_name:<\/em>\n<em>                     getattr(model, output_model_name).trainable = True<\/em>\n<em>                 else:<\/em>\n<em>                     getattr(model, output_model_name).trainable = False<\/em>\n<em>             y<\/em>, loss, grad = train_one_batch(\n                 model, features, targets, label_imbalance,\n                 loss_weights, train_output_model_name, encoder_loss_signs\n             )\n             optimizer.apply_gradients(zip(grad, model.trainable_variables))\n             accuracy[train_output_model_name](\n                 targets[train_output_model_index],\n                 y_[train_output_model_index],\n                 sample_weight=tf.expand_dims(\n                     targets[train_output_model_index]<em>(<\/em>\n<em>                         (<\/em>\n<em>                             (1.0 - label_imbalance[train_output_model_index]) \/ label_imbalance[train_output_model_index]<\/em>\n<em>                         ) -1.0 ) + 1.0, 1<\/em>\n<em>                 )<\/em>\n<em>             )<\/em>\n<em>         epoch_loss_avg(loss)<\/em>\n<em>     print(<\/em>\n<em>         \"{}Epoch {:03d}: Loss: {:.3f}\".format(<\/em>\n<em>             \"<\/em>\" if is_adversarial_emb_training else \" \",\n             epoch,\n             epoch_loss_avg.result()\n         ),\n         *list(map(lambda x: \"{}_accuracy: {:.3f}\".format(x[0], x[1].result()), accuracy.items()))\n     )\n \"\"\"\n *Epoch 000: Loss: -1.209 salary_accuracy: 0.806 sex_accuracy: 0.612 race_accuracy: 0.726\n  Epoch 001: Loss: -0.499 salary_accuracy: 0.812 sex_accuracy: 0.644 race_accuracy: 0.643\n  \u2026\n \"\"\"<\/code><\/pre>\n\n\n\n<p>\u867d\u71363\u4e2a\u795e\u7ecf\u7f51\u7edc\u5934\u7684\u76ee\u6807\u662f\u5c06\u4e2a\u4eba\u5206\u7c7b\u5230\u5c3d\u53ef\u80fd\u9ad8\u7684\u7cbe\u5ea6\uff0c\u4f46\u7f16\u7801\u5668\u7684\u76ee\u6807\u662f\u63d0\u9ad8\u5de5\u8d44\u5206\u7c7b\u5668\u7684\u7cbe\u5ea6\uff0c\u540c\u65f6<strong>\u6d88\u9664\u6027\u522b\u548c\u79cd\u65cf\u7684\u53ef\u9884\u6d4b\u6027<\/strong>\u3002<\/p>\n\n\n\n<p>\u66f4\u4e3a\u590d\u6742\u7684\u662f\uff0c\u6211\u4eec\u8fd8\u5fc5\u987b\u89e3\u51b3\u516c\u5e73\u5b9a\u4e49\u4e2d\u6570\u636e\u96c6\u4e2d\u4e0d\u540c\u7fa4\u4f53\u7684<strong>\u4e25\u91cd\u5931\u8861<\/strong>\u95ee\u9898\u3002\u4e0d\u5e73\u8861\u6570\u636e\u662f\u73b0\u5b9e\u4e16\u754c\u6570\u636e\u96c6\u7684\u4e00\u4e2a\u5e38\u89c1\u6311\u6218\uff0c\u6211\u4eec\u89e3\u51b3\u4e86\u6a21\u578b\u635f\u5931\u548c\u7cbe\u5ea6\u8ba1\u7b97\u4e2d\u7684\u4e0d\u5e73\u8861\u95ee\u9898\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u6211\u4eec\u4f7f\u7528\u81ea\u5b9a\u4e49\u6743\u91cd\u6765\u5bf9\u5e94\u6570\u636e\u7684\u4e0d\u5747\u5300\u6027\uff0c\u5e76\u4f7f\u7528 <em>tf.nn.weighted_cross_entropy_with_logits()<\/em>\u5728\u795e\u7ecf\u7f51\u7edc\u8f93\u51fa\u5c42\u7528\u7ebf\u6027\u6fc0\u6d3b\u51fd\u6570\u6765\u68c0\u7d22logits\u3002<\/p>\n\n\n\n<p>\u6a21\u578b\u5bf9\u6297\u6027\u8bad\u7ec3\u7684\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/miro.medium.com\/max\/1528\/0*OgeTLTnAChOjCE6C\" alt=\"Image for post\"\/><\/figure>\n\n\n\n<p id=\"bac6\">\u6211\u4eec\u6210\u529f\u5730\u4ece\u7f16\u7801\u5668\u4ea7\u751f\u7684\u5171\u4eab\u6570\u636e\u5d4c\u5165\u4e2d\u53bb\u9664\u4e86\u79cd\u65cf\u548c\u6027\u522b\u7684\u53ef\u9884\u6d4b\u6027\u3002\u4f46\u662f\uff0c\u7531\u4e8e\u5168\u7403\u5b58\u5728\u504f\u89c1\uff0c\u85aa\u8d44\u5206\u7c7b\u7684\u51c6\u786e\u6027\u4e5f\u53d7\u5230\u4e86\u4e25\u91cd\u5f71\u54cd\uff0c\u7279\u522b\u662f\u9ad8\u6536\u5165\u8005\u7684\u53ef\u9884\u6d4b\u6027\u3002\u6211\u60f3\u6ca1\u6709\u4ec0\u4e48\u610f\u5916\u3002<\/p>\n\n\n\n<h1 id=\"d93f\">\u5c0f\u7ed3\u4e0e\u6570\u636e\u4f26\u7406<\/h1>\n\n\n\n<p>\u6570\u636e\u79d1\u5b66\u6709\u80fd\u529b\u6539\u53d8\u4f01\u4e1a\uff0c\u5728\u964d\u4f4eB2B\u7684\u98ce\u9669\u548c\u6210\u672c\uff0c\u63d0\u4f9b\u521b\u65b0\u7684B2C\u65b0\u4ea7\u54c1\u65b9\u9762\u505a\u5f97\u5f88\u597d\uff0c\u4f46\u6570\u636e\u79d1\u5b66\u5e94\u7528\u4e5f\u6709\u5f88\u5927\u7684\u8d23\u4efb\u3002<strong>\u6211\u4eec\u5fc5\u987b\u4e0d\u60dc\u4e00\u5207\u4ee3\u4ef7\u907f\u514d\u6570\u636e\u79d1\u5b66\u4ee5\u4e00\u79cd\u4e0d\u53ef\u89e3\u91ca\u548c\u4e0d\u53ef\u89e3\u91ca\u7684\u65b9\u5f0f\u81ea\u52a8\u5f3a\u5316\u504f\u89c1\u3002<\/strong><\/p>\n\n\n\n<p>\u8fd9\u7bc7\u6587\u7ae0\u6709\u5e0c\u671b\u8bf4\u660e\u4e00\u4e9b\u8981\u70b9\uff1a<\/p>\n\n\n\n<ul><li>\u516c\u5e73\u548c\u51cf\u5c11\u504f\u89c1\u662f\u4e00\u4e2a\u590d\u6742\u7684\u8bdd\u9898\uff0c\u6211\u5f53\u7136\u4e0d\u662f\u4e13\u5bb6<\/li><li>\u5728\u6211\u4eec\u751f\u6d3b\u7684\u8fd9\u4e2a\u6709\u504f\u89c1\u7684\u4e16\u754c\u91cc\uff0c\u51cf\u5c11\u504f\u89c1\u4f1a\u5f71\u54cd\u6a21\u578b\u7684\u51c6\u786e\u6027\uff0c\u6211\u4eec\u4f1a\u770b\u5230\u4e00\u4e2a\u51c6\u786e\u6027\u4e0e\u516c\u5e73\u7684\u6743\u8861\u56f0\u5883\u3002<\/li><li>\u4ece\u6a21\u578b\u8f93\u5165\u4e2d\u5220\u9664\u654f\u611f\u6570\u636e\u4e0d\u4f1a\u4f7f\u6a21\u578b\u7ec4\u4e0d\u77e5\u9053\u6216\u4e0d\u516c\u5e73\u3002<\/li><li>\u6211\u4eec\u5fc5\u987b\u9996\u5148\u8bb0\u5f55\u654f\u611f\u6570\u636e\uff0c\u624d\u80fd\u5c06\u5176\u7528\u4e8e\u6211\u4eec\u6a21\u578b\u7684\u5bf9\u6297\u6027\u8bad\u7ec3\u3002\u8fd9\u5bf9\u4e8e\u89e3\u91ca\u6570\u636e\u4fdd\u62a4\u548c\u6570\u636e\u4f26\u7406\u201c\u4f7f\u7528\u4e0e\u7528\u6237\u9700\u6c42\u76f8\u79f0\u7684\u6570\u636e\u201d\u5177\u6709\u91cd\u8981\u610f\u4e49\u3002\u6570\u636e\u4fdd\u62a4\u6cd5\u5bf9\u4e2a\u4eba\u6570\u636e\u975e\u5e38\u660e\u786e\uff0c\u5bf9\u4e2a\u4eba\u53ef\u8bc6\u522b\u7684\u654f\u611f\u6570\u636e\u66f4\u4e3a\u4e25\u683c\u3002\u4e0d\u8bb0\u5f55\u8fd9\u6837\u7684\u6570\u636e\u4f3c\u4e4e\u662f\u6700\u5bb9\u6613\u548c\u6700\u5b89\u5168\u7684\u9009\u62e9\uff0c\u4ee5\u9075\u5b88\u548c\u8868\u9762\u4e0a\u4f3c\u4e4e\u4fc3\u8fdb\u516c\u5e73\u901a\u8fc7\u7fa4\u4f53\u7684\u4e0d\u77e5\u60c5\u3002\u4f46\u6211\u4eec\u5728\u4e0a\u9762\u7684\u7f8e\u56fd\u4eba\u53e3\u666e\u67e5\u6570\u636e\u4f8b\u5b50\u4e2d\u770b\u5230\uff0c\u8fd9\u662f\u9519\u8bef\u7684\uff01\u5b83\u53ef\u80fd\u7b26\u5408\u6570\u636e\u4fdd\u62a4\u6cd5\uff0c\u4f46\u5e76\u6ca1\u6709\u5b9e\u73b0\u516c\u5e73\u3002\u867d\u7136\u6570\u636e\u4fdd\u62a4\u4fc3\u4f7f\u4f01\u4e1a\u4e0d\u6536\u96c6\u654f\u611f\u6570\u636e\uff0c\u4f46\u5408\u4e4e\u9053\u5fb7\u7684\u6570\u636e\u4f7f\u7528\u5b9e\u9645\u4e0a\u6709\u7406\u7531\u8f6c\u800c\u6295\u8d44\u4e8e\u5b89\u5168\u548c\u5408\u89c4\u7684\u654f\u611f\u6570\u636e\u6536\u96c6\u3002<\/li><\/ul>\n\n\n\n<p>\u6b63\u662f\u5728\u8fd9\u4e00\u70b9\u4e0a\uff0c\u9ad8\u7ea7\u4e1a\u52a1\u4e3b\u7ba1\u5fc5\u987b\u627f\u62c5\u4ee5\u4e0b\u8d23\u4efb\uff1a<\/p>\n\n\n\n<ul><li>\u6743\u8861\u6a21\u578b\u7684\u51c6\u786e\u6027\u4e0e\u516c\u5e73\u6027\uff0c\u8fd9\u662f\u4e00\u4e2a\u5f71\u54cd\u76c8\u5229\u7684\u51b3\u7b56<\/li><li>\u662f\u8981\u907f\u514d\u6536\u96c6\u654f\u611f\u6570\u636e\u4ee5\u5b9e\u73b0\u7406\u60f3\u7684\u6570\u636e\u4fdd\u62a4\uff0c\u8fd8\u662f\u6536\u96c6\u654f\u611f\u4fe1\u606f\u4ee5\u4fbf\u5728\u516c\u5e73\u7684\u57fa\u4e8e\u501f\u65b9\u7684\u6a21\u578b\u4e2d\u4ee5\u5408\u4e4e\u9053\u5fb7\u7684\u65b9\u5f0f\u4f7f\u7528\u6570\u636e\uff0c\u8fd9\u662f\u4e00\u4e2a\u5177\u6709\u98ce\u9669\u548c\u6210\u672c\u5f71\u54cd\u7684\u51b3\u7b56\uff08\u4f8b\u5982\u6570\u636e\u6cc4\u9732\u3001\u5b89\u5168\u7684\u6570\u636e\u57fa\u7840\u8bbe\u65bd\u7b49\uff09<\/li><\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u7ffb\u8bd1\u81ea\uff1ahttps:\/\/towardsdatascience.com\/bias-and-algorithm&hellip;&nbsp;<a href=\"https:\/\/xg1990.com\/blog\/archives\/813\" class=\"\" rel=\"bookmark\">\u9605\u8bfb\u66f4\u591a &raquo;<span class=\"screen-reader-text\">\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 &#8211; Bias and Algorithmic Fairness<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"neve_meta_sidebar":"","neve_meta_container":"","neve_meta_enable_content_width":"","neve_meta_content_width":0,"neve_meta_title_alignment":"","neve_meta_author_avatar":"","neve_post_elements_order":"","neve_meta_disable_header":"","neve_meta_disable_footer":"","neve_meta_disable_title":""},"categories":[1],"tags":[],"translation":{"provider":"WPGlobus","version":"3.0.1","language":"zh","enabled_languages":["zh","en"],"languages":{"zh":{"title":true,"content":true,"excerpt":false},"en":{"title":false,"content":false,"excerpt":false}}},"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v18.4.1 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 - Bias and Algorithmic Fairness - xg1990<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/xg1990.com\/blog\/archives\/813\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 - Bias and Algorithmic Fairness - xg1990\" \/>\n<meta property=\"og:url\" content=\"https:\/\/xg1990.com\/blog\/archives\/813\" \/>\n<meta property=\"og:site_name\" content=\"xg1990\" \/>\n<meta property=\"article:published_time\" content=\"2021-02-04T12:42:17+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2021-02-04T12:42:19+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"xg1990\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"5\u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebSite\",\"@id\":\"https:\/\/xg1990.com\/blog\/#website\",\"url\":\"https:\/\/xg1990.com\/blog\/\",\"name\":\"xg1990\",\"description\":\"\u5730\u7406\u5708\u7684\u7801\u519c\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/xg1990.com\/blog\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-CN\"},{\"@type\":\"ImageObject\",\"@id\":\"https:\/\/xg1990.com\/blog\/archives\/813#primaryimage\",\"inLanguage\":\"zh-CN\",\"url\":\"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu\",\"contentUrl\":\"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu\"},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/xg1990.com\/blog\/archives\/813#webpage\",\"url\":\"https:\/\/xg1990.com\/blog\/archives\/813\",\"name\":\"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 - Bias and Algorithmic Fairness - xg1990\",\"isPartOf\":{\"@id\":\"https:\/\/xg1990.com\/blog\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/xg1990.com\/blog\/archives\/813#primaryimage\"},\"datePublished\":\"2021-02-04T12:42:17+00:00\",\"dateModified\":\"2021-02-04T12:42:19+00:00\",\"author\":{\"@id\":\"https:\/\/xg1990.com\/blog\/#\/schema\/person\/774c05a86b0a61bb8ff633749c59fa8f\"},\"breadcrumb\":{\"@id\":\"https:\/\/xg1990.com\/blog\/archives\/813#breadcrumb\"},\"inLanguage\":\"zh-CN\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/xg1990.com\/blog\/archives\/813\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/xg1990.com\/blog\/archives\/813#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\/\/xg1990.com\/blog\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 &#8211; Bias and Algorithmic Fairness\"}]},{\"@type\":\"Person\",\"@id\":\"https:\/\/xg1990.com\/blog\/#\/schema\/person\/774c05a86b0a61bb8ff633749c59fa8f\",\"name\":\"xg1990\",\"image\":{\"@type\":\"ImageObject\",\"@id\":\"https:\/\/xg1990.com\/blog\/#personlogo\",\"inLanguage\":\"zh-CN\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/2d326d725f277a38a498575f23200141?s=96&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/2d326d725f277a38a498575f23200141?s=96&d=mm&r=g\",\"caption\":\"xg1990\"},\"url\":\"https:\/\/xg1990.com\/blog\/archives\/author\/xg1990\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 - Bias and Algorithmic Fairness - xg1990","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/xg1990.com\/blog\/archives\/813","og_locale":"zh_CN","og_type":"article","og_title":"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 - Bias and Algorithmic Fairness - xg1990","og_url":"https:\/\/xg1990.com\/blog\/archives\/813","og_site_name":"xg1990","article_published_time":"2021-02-04T12:42:17+00:00","article_modified_time":"2021-02-04T12:42:19+00:00","og_image":[{"url":"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu"}],"twitter_misc":{"\u4f5c\u8005":"xg1990","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"5\u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebSite","@id":"https:\/\/xg1990.com\/blog\/#website","url":"https:\/\/xg1990.com\/blog\/","name":"xg1990","description":"\u5730\u7406\u5708\u7684\u7801\u519c","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/xg1990.com\/blog\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-CN"},{"@type":"ImageObject","@id":"https:\/\/xg1990.com\/blog\/archives\/813#primaryimage","inLanguage":"zh-CN","url":"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu","contentUrl":"https:\/\/miro.medium.com\/max\/461\/0*2CkXNzgFh5Nsxrsu"},{"@type":"WebPage","@id":"https:\/\/xg1990.com\/blog\/archives\/813#webpage","url":"https:\/\/xg1990.com\/blog\/archives\/813","name":"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 - Bias and Algorithmic Fairness - xg1990","isPartOf":{"@id":"https:\/\/xg1990.com\/blog\/#website"},"primaryImageOfPage":{"@id":"https:\/\/xg1990.com\/blog\/archives\/813#primaryimage"},"datePublished":"2021-02-04T12:42:17+00:00","dateModified":"2021-02-04T12:42:19+00:00","author":{"@id":"https:\/\/xg1990.com\/blog\/#\/schema\/person\/774c05a86b0a61bb8ff633749c59fa8f"},"breadcrumb":{"@id":"https:\/\/xg1990.com\/blog\/archives\/813#breadcrumb"},"inLanguage":"zh-CN","potentialAction":[{"@type":"ReadAction","target":["https:\/\/xg1990.com\/blog\/archives\/813"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/xg1990.com\/blog\/archives\/813#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/xg1990.com\/blog\/"},{"@type":"ListItem","position":2,"name":"\u504f\u89c1\u4e0e\u7b97\u6cd5\u516c\u5e73 &#8211; Bias and Algorithmic Fairness"}]},{"@type":"Person","@id":"https:\/\/xg1990.com\/blog\/#\/schema\/person\/774c05a86b0a61bb8ff633749c59fa8f","name":"xg1990","image":{"@type":"ImageObject","@id":"https:\/\/xg1990.com\/blog\/#personlogo","inLanguage":"zh-CN","url":"https:\/\/secure.gravatar.com\/avatar\/2d326d725f277a38a498575f23200141?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/2d326d725f277a38a498575f23200141?s=96&d=mm&r=g","caption":"xg1990"},"url":"https:\/\/xg1990.com\/blog\/archives\/author\/xg1990"}]}},"_links":{"self":[{"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/posts\/813"}],"collection":[{"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/comments?post=813"}],"version-history":[{"count":2,"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/posts\/813\/revisions"}],"predecessor-version":[{"id":815,"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/posts\/813\/revisions\/815"}],"wp:attachment":[{"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/media?parent=813"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/categories?post=813"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/xg1990.com\/blog\/wp-json\/wp\/v2\/tags?post=813"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}