{"id":1583,"date":"2012-06-03T14:14:58","date_gmt":"2012-06-03T06:14:58","guid":{"rendered":"http:\/\/www.yewen.us\/blog\/?p=1583"},"modified":"2012-06-03T14:14:58","modified_gmt":"2012-06-03T06:14:58","slug":"machine-learning-2-offline-evaluation","status":"publish","type":"post","link":"https:\/\/www.yewen.us\/blog\/2012\/06\/machine-learning-2-offline-evaluation\/","title":{"rendered":"\u673a\u5668\u5b66\u4e60\u624b\u8bb0\u7cfb\u5217 2: \u79bb\u7ebf\u6548\u679c\u8bc4\u4f30"},"content":{"rendered":"

\u4e0a\u4e00\u6b21\u8bf4\u5230\u9009\u7279\u5f81\u7684\u4e00\u4e2a\u7b80\u5355\u65b9\u6cd5, \u4f46\u662f\u5982\u679c\u771f\u7684\u8981\u8bc4\u4f30\u4e00\u4e2a\u65b9\u6cd5\u6216\u8005\u4e00\u7c7b\u7279\u5f81\u7684\u6548\u679c, \u7b80\u5355\u7684\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u662f\u4e0d\u591f\u7684, \u5728\u4e0a\u7ebf\u5b9e\u9a8c\u4e4b\u524d, \u8fd8\u662f\u9700\u8981\u6709\u4e00\u4e9b\u522b\u7684\u65b9\u5f0f\u6765\u505a\u9a8c\u8bc1.<\/p>\n

\u6211\u9047\u5230\u8fc7\u7684\u5927\u90e8\u5206\u673a\u5668\u5b66\u4e60\u95ee\u9898, \u6700\u7ec8\u90fd\u8f6c\u6210\u4e86\u4e8c\u5206\u7c7b\u95ee\u9898 (\u6982\u7387\u95ee\u9898). \u6700\u76f4\u767d\u7684\u6bd4\u5982 A \u662f\u5426\u5c5e\u4e8e\u96c6\u5408 S (\u67d0\u7167\u7247\u4e2d\u7684\u4eba\u8138\u662f\u5426\u662f\u4eba\u7269 Z), \u6392\u5e8f\u95ee\u9898\u4e5f\u53ef\u4ee5\u8f6c\u6362\u4e3a\u4e8c\u5206\u7c7b\u95ee\u9898, \u6bd4\u5982\u5e7f\u544a\u70b9\u51fb\u7387\u6216\u63a8\u8350\u7684\u76f8\u5173\u5ea6, \u628a\u5019\u9009\u96c6\u5206\u4e3a\u70b9\u51fb\/\u4e0d\u70b9\u51fb\u6216\u63a5\u53d7\u63a8\u8350\/\u4e0d\u63a5\u53d7\u63a8\u8350\u7684\u4e8c\u5206\u7c7b\u6982\u7387. \u90a3\u5728\u4e0a\u7ebf\u4e4b\u524d, \u53ef\u4ee5\u7528\u8fc7\u4e00\u4e9b\u5206\u7c7b\u5668\u6027\u80fd\u8bc4\u4f30\u7684\u65b9\u6cd5\u6765\u505a\u79bb\u7ebf\u8bc4\u4f30.<\/p>\n

\u5206\u7c7b\u5668\u7684\u6b63\u786e\u7387\u548c\u53ec\u56de\u7387<\/strong><\/p>\n

\u524d\u51e0\u5929\u5728\u65e0\u89c5\u4e0a\u770b\u5230\u6709\u4eba\u5206\u4eab\u4e86\u4e00\u7bc7 \u6570\u636e\u4e0d\u5e73\u8861\u65f6\u5206\u7c7b\u5668\u6027\u80fd\u8bc4\u4ef7\u4e4bROC\u66f2\u7ebf\u5206\u6790<\/a>, \u628a\u8fd9\u4e2a\u95ee\u9898\u5df2\u7ecf\u8bb2\u5dee\u4e0d\u591a\u4e86, \u6211\u8fd9\u590d\u8ff0\u4e00\u4e0b.<\/p>\n

\u5148\u8bf4\u6df7\u6dc6\u77e9\u9635 (confusion matrix). \u6df7\u6dc6\u77e9\u9635\u662f\u8bc4\u4f30\u5206\u7c7b\u5668\u53ef\u4fe1\u5ea6\u7684\u4e00\u4e2a\u57fa\u672c\u5de5\u5177, \u8bbe\u5b9e\u9645\u7684\u6240\u6709\u6b63\u6837\u672c\u4e3a P (real-Positive), \u8d1f\u6837\u672c\u4e3a N (real-Negative), \u5206\u7c7b\u5668\u5206\u5230\u7684\u6b63\u6837\u672c\u6807\u4e3a pre-Positive’, \u8d1f\u6837\u672c\u6807\u4e3a pre-Negetive’, \u5219\u53ef\u4ee5\u7528\u4e0b\u9762\u7684\u6df7\u6dc6\u77e9\u9635\u8868\u793a\u6240\u6709\u60c5\u51b5:<\/p>\n

              | real-positive       | real-negative\npre-positive' | TP (true positive)  | FP (false positive)\npre-negative' | FN (false negative) | TN (true negative)<\/pre>\n
\u901a\u8fc7\u8fd9\u4e2a\u77e9\u9635, \u53ef\u4ee5\u5f97\u5230\u5f88\u591a\u8bc4\u4f30\u6307\u6807:<\/p>\n
FP rate = FP \/ N\nTP rate = TP \/ P\nAccuracy = (TP + TN) \/ (P + N)    # \u4e00\u822c\u79f0\u4e4b\u4e3a\u51c6\u786e\u6027\u6216\u6b63\u786e\u6027\nPrecision = TP \/ (TP + FP)        # \u53e6\u4e00\u4e9b\u9886\u57df\u7684\u51c6\u786e\u6027\u6216\u6b63\u786e\u6027, \u7ecf\u5e38\u9700\u8981\u770b\u4e0a\u4e0b\u6587\u6765\u5224\u65ad\nRecall = TP \/ P                   # \u4e00\u822c\u79f0\u4e4b\u4e3a\u53ec\u56de\u7387\nF-score = Precision * Recall<\/pre>\n
\u5728\u6211\u63a5\u89e6\u8fc7\u7684\u5927\u90e8\u5206\u5de5\u4f5c\u4e2d, \u5927\u5bb6\u90fd\u5728\u5173\u6ce8 Precision \u548c Recall. \u540c\u5f15\u7528\u539f\u6587\u4e2d\u63d0\u5230\u7684, \u8fd9\u6837\u7684\u5206\u7c7b\u8bc4\u4f30\u6027\u80fd\u53ea\u5728\u6570\u636e\u6bd4\u8f83\u5e73\u8861\u65f6\u6bd4\u8f83\u597d\u7528 (\u6b63\u8d1f\u4f8b\u6bd4\u4f8b\u63a5\u8fd1), \u5728\u5f88\u591a\u7279\u5b9a\u60c5\u51b5\u4e0b\u6b63\u8d1f\u4f8b\u662f\u660e\u663e\u6709\u504f\u7684 (\u6bd4\u5982\u4e07\u5206\u4e4b\u51e0\u70b9\u51fb\u7387\u7684\u663e\u793a\u5e7f\u544a), \u90a3\u5c31\u53ea\u80fd\u4f5c\u4e3a\u4e00\u5b9a\u7684\u53c2\u8003\u6307\u6807.<\/p>\n
\u5206\u7c7b\u5668\u7684\u6392\u5e8f\u80fd\u529b\u8bc4\u4f30<\/strong><\/p>\n
\u5f88\u591a\u60c5\u51b5\u4e0b\u6211\u4eec\u9664\u4e86\u5e0c\u671b\u5206\u7c7b\u5668\u6309\u67d0\u4e2a\u9608\u503c\u5c06\u6b63\u8d1f\u6837\u672c\u5b8c\u5168\u5206\u5f00, \u540c\u65f6\u8fd8\u60f3\u77e5\u9053\u5019\u9009\u96c6\u4e2d\u4e0d\u540c\u6761\u76ee\u7684\u5e8f\u5173\u7cfb. \u6bd4\u5982\u5e7f\u544a\u548c\u63a8\u8350, \u9996\u5148\u9700\u8981\u4e00\u4e2a\u57fa\u7840\u9608\u503c\u6765\u4fdd\u8bc1\u53ec\u56de\u7684\u5185\u5bb9\u90fd\u6ee1\u8db3\u57fa\u672c\u76f8\u5173\u5ea6, \u6bd4\u5982\u6211\u4e00\u5927\u8001\u7237\u4eec\u53bb\u641c\u7b14\u8bb0\u672c\u7ef4\u4fee\u4ee3\u7406\u4f60\u7ed9\u6211\u51fa\u4e00\u5c11\u5973\u776b\u6bdb\u818f\u7684\u5e7f\u544a\u6216\u63a8\u8350\u5173\u6ce8, \u6211\u7edd\u5bf9\u98d9\u4e00\u53e5\u4f60\u5927\u7237\u7684\u7136\u540e\u5f00 AdBlock<\/a> \u5c4f\u853d\u4e4b. \u5728\u4fdd\u8bc1\u4e86\u57fa\u7840\u76f8\u5173\u6027 (\u5373\u5206\u7c7b\u5668\u7684\u6b63\u8d1f\u4f8b\u5206\u5f00) \u540e, \u5219\u9700\u8981\u6bd4\u8f83\u540c\u6837\u662f\u6b63\u4f8b\u7684\u96c6\u5408\u91cc, \u54ea\u4e9b\u66f4\u6b63\u70b9 (\u5176\u5b9e\u8bf4\u767d\u4e86\u5c31\u662f\u600e\u6837\u624d\u6536\u76ca\u6700\u5927\u5316). \u4e00\u822c\u6765\u8bf4, \u5982\u679c\u5206\u7c7b\u5668\u7684\u8f93\u51fa\u662f\u4e00\u4e2a\u6b63\u4f8b\u6982\u7387, \u5219\u76f4\u63a5\u6309\u8fd9\u4e2a\u6982\u7387\u6765\u6392\u5e8f\u5c31\u884c\u4e86. \u5982\u679c\u6700\u7ec8\u6536\u76ca\u8fd8\u8981\u901a\u8fc7\u8bc4\u4f30\u51fd\u6570\u8f6c\u6362, \u6bd4\u5982\u5e7f\u544a\u7684 eCPM = CTR*Price, \u6216\u63a8\u8350\u91cc rev = f(CTR), (f(x) \u662f\u4e00\u4e2a\u4e0d\u540c\u6761\u76ee\u7684\u83b7\u76ca\u6743\u91cd\u51fd\u6570), \u90a3\u4e48\u4e3a\u4e86\u8bc4\u4f30\u5e8f\u662f\u5426\u597d, \u4e00\u822c\u4f1a\u518d\u5f15\u5165 ROC \u66f2\u7ebf\u548c AUC \u9762\u79ef\u4e24\u4e2a\u6307\u6807.<\/p>\n
ROC \u66f2\u7ebf\u5168\u79f0\u662f Receiver Operating Characteristic (ROC curve), \u8be6\u7ec6\u7684\u89e3\u91ca\u53ef\u4ee5\u89c1\u7ef4\u57fa\u767e\u79d1\u4e0a\u7684\u82f1\u6587\u8bcd\u6761 Receiver_operating_characteristic<\/a> \u6216\u4e2d\u6587\u8bcd\u6761 ROC\u66f2\u7ebf<\/a>. \u6211\u5bf9 ROC \u66f2\u7ebf\u7684\u7406\u89e3\u662f, \u5bf9\u67d0\u4e2a\u6837\u672c\u96c6, \u5f53\u524d\u5206\u7c7b\u5668\u5bf9\u5176\u5206\u7c7b\u7ed3\u679c\u7684 FPR \u5728 x \u65f6, TPR \u80fd\u5230 y. \u5982\u679c\u5206\u7c7b\u5668\u5b8c\u5168\u51c6\u786e, \u5219\u5728 x = 0 \u65f6 y \u5c31\u80fd\u5230 1, \u5982\u679c\u5206\u7c7b\u5668\u5b8c\u5168\u4e0d\u9760\u8c31, \u5219\u5728 x = 1 \u65f6 y \u8fd8\u662f\u4e3a 0, \u5982\u679c x = y, \u90a3\u8bf4\u660e\u8fd9\u4e2a\u5206\u7c7b\u5668\u5728\u968f\u673a\u5206\u7c7b. \u56e0\u4e3a\u4e24\u4e2a\u90fd\u662f Rate, \u662f [0, 1] \u4e4b\u95f4\u7684\u53d6\u503c, \u6240\u4ee5\u6309\u6b64\u65b9\u6cd5\u63cf\u7684\u70b9\u90fd\u5728\u4e00\u4e2a (0, 0), (1, 1) \u7684\u77e9\u5f62\u5185, \u62c9\u4e00\u6761\u76f4\u7ebf\u4ece (0, 0) \u5230 (1, 1), \u5982\u679c\u63cf\u70b9\u5728\u8fd9\u6761\u76f4\u7ebf\u4e0a, \u8bf4\u660e\u5206\u7c7b\u5668\u5bf9\u5f53\u524d\u6837\u672c\u5c31\u662f\u968f\u673a\u5206\u7684 (\u505a\u5206\u7c7b\u6700\u60b2\u50ac\u7684\u4e8b), \u5982\u679c\u63cf\u70b9\u5728\u5de6\u4e0a\u65b9, \u8bf4\u660e\u5f53\u524d\u5206\u7c7b\u5668\u5bf9\u6b64\u6837\u672c\u5206\u7c7b\u6548\u679c\u597d\u8fc7\u968f\u673a, \u5982\u679c\u5728\u53f3\u4e0b\u65b9, \u90a3\u8bf4\u660e\u5206\u7c7b\u5668\u5728\u505a\u6bd4\u968f\u673a\u8fd8\u5751\u7239\u7684\u53cd\u5411\u5206\u7c7b. \u5f15\u7528\u7ef4\u57fa\u767e\u79d1\u4e0a\u7684\u4e00\u4e2a\u56fe\u6765\u8bf4\u660e:
\n
\"ROC<\/center>
\n\u5176\u4e2d C’ \u597d\u4e8e A (\u90fd\u662f\u6b63\u5411\u7684), B \u662f\u968f\u673a, C \u662f\u4e00\u4e2a\u53cd\u6548\u679c (\u8ddf C’ \u6cbf\u7ea2\u7ebf\u8f74\u5bf9\u79f0, \u5c31\u662f\u8bf4\u628a C \u7684\u7ed3\u679c\u53cd\u8fc7\u6765\u5c31\u5f97\u5230 C’).<\/p>\n
\u5982\u679c\u6211\u4eec\u6709\u8db3\u591f\u591a\u7684\u6837\u672c, \u5219\u5bf9\u4e00\u4e2a\u5206\u7c7b\u5668\u53ef\u4ee5\u5728 ROC \u66f2\u7ebf\u56fe\u4e0a\u753b\u51fa\u82e5\u5e72\u4e2a\u70b9, \u628a\u8fd9\u4e9b\u70b9\u548c (0, 0), (1, 1) \u8fde\u8d77\u6765\u6c42\u51f8\u5305, \u5c31\u5f97\u5230\u4e86 AUC \u9762\u79ef (Area Under Curve, \u66f2\u7ebf\u4e0b\u9762\u79ef). \u975e\u5e38\u660e\u663e, \u8fd9\u4e2a\u51f8\u5305\u7684\u6700\u5c0f\u4e0b\u9762\u79ef\u662f 0.5 (\u4ece (0, 0) \u5230 (1, 1) \u7684\u8fd9\u6761\u7ebf), \u6700\u5927\u662f 1.0 (\u6574\u4e2a\u77e9\u5f62\u9762\u79ef), AUC \u503c\u8d8a\u5927, \u8bf4\u660e\u5206\u7c7b\u6548\u679c\u8d8a\u597d.<\/p>\n
\u7528 ROC \u66f2\u7ebf\u5b9a\u4e49\u7684\u65b9\u5f0f\u6765\u63cf\u70b9\u8ba1\u7b97\u9762\u79ef\u4f1a\u5f88\u9ebb\u70e6, \u4e0d\u8fc7\u8fd8\u597d\u524d\u4eba\u7ed9\u4e86\u6211\u4eec\u4e00\u4e2a\u8fd1\u4f3c\u516c\u5f0f, \u6211\u627e\u5230\u7684\u6700\u539f\u59cb\u51fa\u5904\u662f Hand, Till \u5728 Machine Learning 2001 \u4e0a\u7684\u4e00\u7bc7\u6587\u7ae0\u7ed9\u51fa [\u6587\u7ae0\u94fe\u63a5<\/a>]. \u4e2d\u95f4\u7684\u63a8\u5bfc\u8fc7\u7a0b\u6bd4\u8f83\u7e41\u7410, \u76f4\u63a5\u8bf4\u6211\u5bf9\u8fd9\u4e2a\u8ba1\u7b97\u65b9\u6cd5\u7684\u7406\u89e3: \u5c06\u6240\u6709\u6837\u672c\u6309\u9884\u4f30\u6982\u7387\u4ece\u5c0f\u5230\u5927\u6392\u5e8f, \u7136\u540e\u4ece (0, 0) \u70b9\u5f00\u59cb\u63cf\u70b9, \u6bcf\u4e2a\u65b0\u7684\u70b9\u662f\u5728\u524d\u4e00\u4e2a\u70b9\u7684\u57fa\u7840\u4e0a, \u6a2a\u5750\u6807\u52a0\u4e0a\u5f53\u524d\u6837\u672c\u7684\u6b63\u4f8b\u5728\u603b\u6b63\u4f8b\u6570\u4e2d\u7684\u5360\u6bd4, \u7eb5\u5750\u6807\u52a0\u4e0a\u5f53\u524d\u6837\u672c\u7684\u8d1f\u4f8b\u5728\u603b\u8d1f\u4f8b\u6570\u4e2d\u7684\u5360\u6bd4, \u6700\u7ec8\u7684\u7ec8\u70b9\u4e00\u5b9a\u662f (1, 1), \u5bf9\u8fd9\u4e2a\u66f2\u7ebf\u6c42\u9762\u79ef, \u5373\u5f97\u5230 AUC. \u5176\u7269\u7406\u610f\u4e49\u4e5f\u975e\u5e38\u76f4\u89c2, \u5982\u679c\u6211\u4eec\u628a\u8d1f\u4f8b\u90fd\u6392\u5728\u6b63\u4f8b\u524d\u9762, \u5219\u66f2\u7ebf\u4e00\u5b9a\u662f\u5148\u5f80\u4e0a\u518d\u5f80\u53f3, \u5f97\u5230\u7684\u9762\u79ef\u5927\u4e8e 0.5, \u8bf4\u660e\u5206\u7c7b\u5668\u6548\u679c\u6bd4\u968f\u673a\u597d, \u6700\u6781\u7aef\u7684\u60c5\u51b5\u5c31\u662f\u6240\u6709\u8d1f\u4f8b\u90fd\u5728\u6b63\u4f8b\u524d, \u5219\u66f2\u7ebf\u5c31\u662f (0, 0) -> (0, 1) -> (1, 1) \u8fd9\u6837\u7684\u5f62\u72b6, \u9762\u79ef\u4e3a 1.0.<\/p>\n
\u540c\u6837\u7ed9\u4e00\u4efd C \u4ee3\u7801\u5b9e\u73b0:<\/p>\n
struct SampleNode {\n  double predict_value;\n  unsigned int pos_num;\n  unsigned int neg_num;\n};\n\nint cmp(const void *a, const void *b)\n{\n   SampleNode *aa = (SampleNode *)a;\n   SampleNode *bb = (SampleNode *)b;\n   return(((aa->predict_value)-(bb->predict_value)>0)?1:-1);\n}\n\ndouble calcAuc(SampleNode samples[], int sample_num) {\n  qsort(samples, sample_num, sizeof(SampleNode), cmp);\n\n  \/\/ init all counters\n  double sum_pos = 0;\n  double sum_neg = 0;\n  double new_neg = 0;\n  double rp = 0;\n  for (int i = 0; i < sample_num; ++i) {\n    if (samples[i].neg_num >= 0) {\n      new_neg += samples[i].neg_num;\n    }\n\n    if (samples[i].pos_num >= 0) {\n      \/\/ calc as trapezium, not rectangle\n      rp += samples[i].pos_num * (sum_neg + new_neg)\/2;\n      sum_pos += samples[i].pos_num;\n    }\n    sum_neg = new_neg;\n  }\n\n  return rp\/(sum_pos*sum_neg);\n}<\/pre>\n
\u5206\u7c7b\u5668\u7684\u4e00\u81f4\u6027<\/strong><\/p>\n
\u5982\u679c\u5206\u7c7b\u5668\u7684\u6982\u7387\u7ed3\u679c\u5c31\u662f\u6700\u7ec8\u7ed3\u679c\u5e8f, \u90a3 AUC \u503c\u57fa\u672c\u53ef\u4ee5\u5f53\u4f5c\u6700\u7ec8\u6548\u679c\u6765\u7528. \u4f46\u662f\u5b9e\u9645\u5e94\u7528\u4e2d\u5206\u7c7b\u5668\u7684\u7ed3\u679c\u90fd\u8981\u518d\u505a\u51fd\u6570\u8f6c\u6362\u624d\u662f\u6700\u7ec8\u5e8f, \u5219\u8bc4\u4f30\u7684\u65f6\u5019\u9700\u8981\u5c06\u8f6c\u6362\u51fd\u6570\u4e5f\u5e26\u4e0a\u53bb\u505a AUC \u8bc4\u4f30\u624d\u884c. \u67d0\u4e9b\u5e94\u7528\u4e2d\u8fd9\u4e2a\u8f6c\u6362\u51fd\u6570\u662f\u4e0d\u786e\u5b9a\u7684, \u6bd4\u5982\u5e7f\u544a\u7684\u4ef7\u683c\u968f\u65f6\u4f1a\u53d8, \u63a8\u8350\u6761\u76ee\u7684\u91cd\u8981\u6027\u6216\u6536\u76ca\u53ef\u80fd\u4e5f\u662f\u53e6\u4e00\u4e2a\u8ba1\u7b97\u6a21\u578b\u7684\u7ed3\u679c. \u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b, \u5982\u679c\u6211\u4eec\u53ef\u4ee5\u4fdd\u8bc1\u5206\u7c7b\u5668\u6982\u7387\u548c\u5b9e\u9645\u6982\u7387\u4e00\u81f4, \u8ba9\u540e\u7eed\u7684\u8f6c\u6362\u51fd\u6570\u62ff\u5230\u4e00\u4e2a\u6b63\u786e\u7684\u8f93\u5165, \u90a3\u4e48\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u624d\u80fd\u8fbe\u5230\u6700\u4f18\u6027\u80fd.<\/p>\n
\u4e3a\u4e86\u8bc4\u4f30\u5206\u7c7b\u5668\u6982\u7387\u548c\u5b9e\u9645\u6982\u7387\u7684\u4e00\u81f4\u6027, \u5f15\u5165 MAE (Mean Absolute Error, \u5e73\u5747\u7edd\u5bf9\u8bef\u5dee) \u8fd9\u4e2a\u6307\u6807, \u7ef4\u57fa\u767e\u79d1\u5bf9\u5e94\u7684\u8bcd\u6761\u662f Mean_absolute_error<\/a>. \u6700\u7ec8\u7684\u8ba1\u7b97\u65b9\u6cd5\u5f88\u7b80\u5355, \u5bf9\u6837\u672c i, fi<\/sub> \u662f\u9884\u4f30\u6982\u7387, yi<\/sub> \u662f\u5b9e\u9645\u6982\u7387, \u5219 i \u4e0a\u7edd\u5bf9\u8bef\u5dee\u662f ei<\/sub>, \u7d2f\u52a0\u6c42\u5e73\u5747\u5c31\u662f MAE:
\n
\"MAE<\/center><\/p>\n
MAE \u7684\u503c\u57df\u662f [0, +∞), \u503c\u8d8a\u5c0f\u8bf4\u660e\u5206\u7c7b\u5668\u8f93\u51fa\u548c\u5b9e\u9645\u503c\u7684\u4e00\u81f4\u6027\u8d8a\u597d. \u6211\u4e2a\u4eba\u8ba4\u4e3a\u5982\u679c MAE \u548c\u5b9e\u9645\u6982\u7387\u4e00\u6837\u5927, \u90a3\u8fd9\u4e2a\u5206\u7c7b\u5668\u7684\u6ce2\u52a8\u6548\u679c\u4e5f\u5927\u5230\u8ba9\u9884\u4f30\u8fd1\u4f3c\u968f\u673a\u4e86.<\/p>\n
MAE \u770b\u8d77\u6765\u548c\u6807\u51c6\u5dee\u6709\u70b9\u50cf, \u7c7b\u4f3c\u6807\u51c6\u5dee\u548c\u65b9\u5dee\u7684\u5173\u7cfb, MAE \u4e5f\u6709\u4e00\u4e2a\u5bf9\u5e94\u7684 MSE (Mean Squared Error, \u5747\u65b9\u5dee?), \u8fd9\u4e2a\u6307\u6807\u66f4\u591a\u8003\u8651\u7684\u662f\u6781\u574f\u60c5\u51b5\u7684\u5f71\u54cd, \u8ba1\u7b97\u6bd4\u8f83\u9ebb\u70e6, \u4e00\u822c\u7528\u7684\u4e5f\u4e0d\u591a, \u6709\u5174\u8da3\u7684\u53ef\u4ee5\u770b\u7ef4\u57fa\u767e\u79d1\u4e0a\u7684\u8bcd\u6761 Mean_squared_error<\/a>.<\/p>\n
MAE \u8ba1\u7b97\u592a\u7b80\u5355, MSE \u8ba1\u7b97\u592a\u7ea0\u7ed3, \u6240\u4ee5\u90fd\u4e0d\u5728\u8fd9\u7ed9\u51fa\u4ee3\u7801\u5b9e\u73b0.<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4e0a\u4e00\u6b21\u8bf4\u5230\u9009\u7279\u5f81\u7684\u4e00\u4e2a\u7b80\u5355\u65b9\u6cd5, \u4f46\u662f\u5982\u679c\u771f\u7684\u8981\u8bc4\u4f30\u4e00\u4e2a\u65b9\u6cd5\u6216\u8005\u4e00\u7c7b\u7279\u5f81\u7684\u6548\u679c, \u7b80\u5355\u7684\u76f8\u4f3c\u5ea6\u8ba1\u7b97\u662f\u4e0d\u591f\u7684,  […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[5,7,9],"tags":[29,101,219,310,327,409],"class_list":["post-1583","post","type-post","status-publish","format-standard","hentry","category-work-life","category-tech-notes","category-machine-learning","tag-auc","tag-mae","tag-219","tag-310","tag-327","tag-409"],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p4aR5e-px","_links":{"self":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts\/1583","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/comments?post=1583"}],"version-history":[{"count":0,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/posts\/1583\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/media?parent=1583"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/categories?post=1583"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.yewen.us\/blog\/wp-json\/wp\/v2\/tags?post=1583"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}