{"id":89949,"date":"2026-08-04T10:00:41","date_gmt":"2026-08-04T02:00:41","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/89949.html"},"modified":"2026-08-04T10:00:41","modified_gmt":"2026-08-04T02:00:41","slug":"%e3%80%90%e6%9c%ba%e5%99%a8%e5%ad%a6%e4%b9%a0%e4%b8%93%e6%a0%8f%e3%80%915-1-%e8%ae%ad%e7%bb%83%e5%b7%a5%e7%a8%8b%ef%bc%9a%e6%95%b0%e6%8d%ae%e4%b8%8d%e5%b9%b3%e8%a1%a1%e5%a4%84%e7%90%86","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/89949.html","title":{"rendered":"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406"},"content":{"rendered":"

\u5f15\u5b50：\u4e09\u9053\u8ba9\u4f60\u6000\u7591\u4eba\u751f\u7684\u9762\u8bd5\u9898<\/h3>\n

\u5728\u5f00\u59cb\u4e4b\u524d，\u8bf7\u5148\u601d\u8003\u4e0b\u9762\u51e0\u4e2a\u95ee\u9898\u3002\u5982\u679c\u4f60\u80fd\u4e00\u53e3\u6c14\u5168\u90e8\u7b54\u51fa\u6765，\u8bf4\u660e\u4f60\u5bf9\u6570\u636e\u4e0d\u5e73\u8861\u5df2\u7ecf\u6709\u4e86\u6bd4\u8f83\u900f\u5f7b\u7684\u7406\u89e3；\u5982\u679c\u7b54\u4e0d\u4e0a\u6765，\u90a3\u4e48\u8fd9\u4e00\u7ae0\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002<\/p>\n

\u95ee\u9898 1：\u4f60\u5728\u4e00\u4e2a\u4e8c\u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3\u6a21\u578b，\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387\u8fbe\u5230\u4e86 99.3%\u3002\u4f60\u975e\u5e38\u9ad8\u5174，\u4f46\u4f60\u7684\u4e0a\u7ea7\u770b\u4e86\u4e00\u773c\u6570\u636e\u5206\u5e03\u8bf4：\u201c\u6b63\u6837\u672c\u53ea\u5360 0.7%，\u4f60\u8fd9\u6a21\u578b\u4ec0\u4e48\u90fd\u6ca1\u5b66\u5230\u3002\u201d \u4e3a\u4ec0\u4e48？\u8be5\u600e\u4e48\u907f\u514d\u8fd9\u79cd\u60c5\u51b5？<\/p>\n

\u95ee\u9898 2：\u9762\u8bd5\u5b98\u8ba9\u4f60\u624b\u5199 Focal Loss \u7684 PyTorch \u5b9e\u73b0\u3002\u4f60\u5199\u51fa\u6765\u4e86，\u4f46\u4ed6\u8ffd\u95ee：\u201c\u4e3a\u4ec0\u4e48 Focal Loss \u80fd\u7f13\u89e3\u4e0d\u5e73\u8861？\u8c03\u8282\u56e0\u5b50 gamma \u8d77\u5230\u4ec0\u4e48\u4f5c\u7528？\u201d \u4f60\u80fd\u4ece\u68af\u5ea6\u89d2\u5ea6\u89e3\u91ca\u6e05\u695a\u5417？<\/p>\n

\u95ee\u9898 3：\u4f60\u7684\u6570\u636e\u4e2d\u6b63\u8d1f\u6837\u672c\u6bd4\u4f8b\u662f 1:10,000\u3002\u4f60\u8bd5\u4e86 SMOTE，\u6a21\u578b\u53cd\u800c\u53d8\u5dee\u4e86\u3002\u9762\u8bd5\u5b98\u8bf4：\u201cSMOTE \u5728\u9ad8\u4e0d\u5e73\u8861\u7387\u4e0b\u53ef\u80fd\u4f1a\u5f15\u5165\u566a\u58f0\u3002\u4f60\u8bd5\u8bd5\u5176\u4ed6\u65b9\u6cd5\u3002\u201d \u4f60\u80fd\u8bf4\u51fa\u81f3\u5c11\u4e09\u79cd\u6539\u8fdb\u65b9\u6848\u5e76\u8bf4\u660e\u5b83\u4eec\u5404\u81ea\u7684\u9002\u7528\u573a\u666f\u5417？<\/p>\n

\n

2.1 \u51c6\u786e\u7387\u9677\u9631 \u2014\u2014 99% \u51c6\u786e\u7387\u7684"\u865a\u5047\u7e41\u8363"<\/h3>\n

2.1.1 \u4e3a\u4ec0\u4e48\u51c6\u786e\u7387\u4f1a\u9a97\u4eba？<\/h4>\n

\u51c6\u786e\u7387（Accuracy）\u7684\u5b9a\u4e49\u662f：<\/p>\n

Accuracy=TP+TNTP+TN+FP+FN \\\\text{Accuracy} = \\\\frac{TP + TN}{TP + TN + FP + FN} <\/span><\/span>Accuracy<\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>T<\/span>P<\/span><\/span>+<\/span><\/span>T<\/span>N<\/span><\/span>+<\/span><\/span>F<\/span>P<\/span><\/span>+<\/span><\/span>F<\/span>N<\/span><\/span><\/span><\/span><\/span><\/span><\/span>T<\/span>P<\/span><\/span>+<\/span><\/span>T<\/span>N<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u5f53\u7c7b\u522b\u6781\u5ea6\u4e0d\u5e73\u8861\u65f6，\u8fd9\u4e2a\u6307\u6807\u4f1a\u5b8c\u5168"\u5931\u7075"\u3002\u8003\u8651\u4e00\u4e2a\u6b3a\u8bc8\u68c0\u6d4b\u573a\u666f：100,000 \u7b14\u4ea4\u6613\u4e2d\u53ea\u6709 100 \u7b14\u662f\u6b3a\u8bc8（\u6b63\u6837\u672c\u7387 = 0.1%）\u3002\u5982\u679c\u4e00\u4e2a\u6a21\u578b\u628a\u6240\u6709\u4ea4\u6613\u90fd\u5224\u4e3a"\u6b63\u5e38"：<\/p>\n

Accuracy=0+99,\u2009\u2063900100,\u2009\u2063000=99.9% \\\\text{Accuracy} = \\\\frac{0 + 99,\\\\!900}{100,\\\\!000} = 99.9\\\\% <\/span><\/span>Accuracy<\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>100<\/span>,<\/span><\/span><\/span>000<\/span><\/span><\/span><\/span><\/span><\/span><\/span>0<\/span><\/span>+<\/span><\/span>99<\/span>,<\/span><\/span><\/span>900<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>99.9%<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u8fd9\u4e2a 99.9% \u7684\u51c6\u786e\u7387\u770b\u4f3c\u5b8c\u7f8e，\u4f46\u6a21\u578b\u5b9e\u9645\u4e0a\u4ec0\u4e48\u90fd\u6ca1\u505a\u2014\u2014\u5b83\u8fde\u4e00\u4e2a\u6b3a\u8bc8\u6837\u672c\u90fd\u6ca1\u8bc6\u522b\u51fa\u6765！<\/p>\n

2.1.2 \u66f4\u53ef\u9760\u7684\u8bc4\u4f30\u6307\u6807<\/h4>\n\n\u6307\u6807\u5b9a\u4e49\u5bf9\u4e0d\u5e73\u8861\u7684\u9c81\u68d2\u6027<\/tr>\n\n\n\n\n\n\n\n\n
\u7cbe\u786e\u7387 (Precision)<\/td>\nTP \/ (TP + FP)<\/td>\n\u4e2d\u7b49<\/td>\n<\/tr>\n
\u53ec\u56de\u7387 (Recall)<\/td>\nTP \/ (TP + FN)<\/td>\n\u4e2d\u7b49<\/td>\n<\/tr>\n
F1-Score<\/td>\n2 x P x R \/ (P + R)<\/td>\n\u8f83\u597d<\/td>\n<\/tr>\n
AUC-ROC<\/td>\nROC \u66f2\u7ebf\u4e0b\u9762\u79ef<\/td>\n\u8f83\u597d<\/td>\n<\/tr>\n
AUC-PR<\/td>\nPrecision-Recall \u66f2\u7ebf\u4e0b\u9762\u79ef<\/td>\n\u6700\u597d<\/td>\n<\/tr>\n
G-Mean<\/td>\nsqrt(Recall x Specificity)<\/td>\n\u8f83\u597d<\/td>\n<\/tr>\n
Matthews \u76f8\u5173\u7cfb\u6570<\/td>\n\u7efc\u5408\u8003\u8651\u56db\u7c7b\u9884\u6d4b\u7ed3\u679c<\/td>\n\u6700\u597d<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u5173\u952e\u8ba4\u77e5：\u5728\u6781\u5ea6\u4e0d\u5e73\u8861\u573a\u666f\u4e0b，AUC-PR \u6bd4 AUC-ROC \u66f4\u654f\u611f\u3002\u56e0\u4e3a ROC \u66f2\u7ebf\u53d7\u8d1f\u6837\u672c（\u591a\u6570\u7c7b）\u5f71\u54cd\u8f83\u5927，\u800c PR \u66f2\u7ebf\u805a\u7126\u4e8e\u6b63\u6837\u672c（\u5c11\u6570\u7c7b）\u7684\u8868\u73b0\u3002<\/p>\n

import<\/span> numpy as<\/span> np
\nfrom<\/span> sklearn.<\/span>metrics import<\/span> (<\/span>
\n accuracy_score,<\/span> precision_score,<\/span> recall_score,<\/span> f1_score,<\/span>
\n roc_auc_score,<\/span> average_precision_score,<\/span> matthews_corrcoef
\n)<\/span><\/p>\n

def<\/span> evaluate_imbalanced<\/span>(<\/span>y_true,<\/span> y_pred,<\/span> y_prob=<\/span>None<\/span>)<\/span>:<\/span>
\n """\u5168\u9762\u8bc4\u4f30\u4e0d\u5e73\u8861\u6570\u636e\u4e0b\u7684\u6a21\u578b\u8868\u73b0"""<\/span>
\n metrics =<\/span> {<\/span>
\n 'accuracy'<\/span>:<\/span> accuracy_score(<\/span>y_true,<\/span> y_pred)<\/span>,<\/span>
\n 'precision'<\/span>:<\/span> precision_score(<\/span>y_true,<\/span> y_pred,<\/span> zero_division=<\/span>0<\/span>)<\/span>,<\/span>
\n 'recall'<\/span>:<\/span> recall_score(<\/span>y_true,<\/span> y_pred,<\/span> zero_division=<\/span>0<\/span>)<\/span>,<\/span>
\n 'f1'<\/span>:<\/span> f1_score(<\/span>y_true,<\/span> y_pred,<\/span> zero_division=<\/span>0<\/span>)<\/span>,<\/span>
\n 'mcc'<\/span>:<\/span> matthews_corrcoef(<\/span>y_true,<\/span> y_pred)<\/span>,<\/span>
\n }<\/span>
\n if<\/span> y_prob is<\/span> not<\/span> None<\/span>:<\/span>
\n metrics[<\/span>'auc_roc'<\/span>]<\/span> =<\/span> roc_auc_score(<\/span>y_true,<\/span> y_prob)<\/span>
\n metrics[<\/span>'auc_pr'<\/span>]<\/span> =<\/span> average_precision_score(<\/span>y_true,<\/span> y_prob)<\/span>
\n return<\/span> metrics<\/p>\n

# \u793a\u4f8b：\u5168\u90e8\u5224\u8d1f\u7684"\u5047\u6a21\u578b"<\/span>
\ny_true =<\/span> np.<\/span>array(<\/span>[<\/span>0<\/span>]<\/span> *<\/span> 99900<\/span> +<\/span> [<\/span>1<\/span>]<\/span> *<\/span> 100<\/span>)<\/span>
\ny_pred =<\/span> np.<\/span>array(<\/span>[<\/span>0<\/span>]<\/span> *<\/span> 100000<\/span>)<\/span> # \u5168\u90e8\u5224\u8d1f<\/span><\/p>\n

print<\/span>(<\/span>evaluate_imbalanced(<\/span>y_true,<\/span> y_pred)<\/span>)<\/span>
\n# \u8f93\u51fa: accuracy=0.999, precision=0.0, recall=0.0, f1=0.0, mcc=0.0<\/span><\/p>\n

\n

\u9762\u8bd5\u5b98\u8ffd\u95ee\u94fe<\/h4>\n

Q：\u9664\u4e86\u6362\u6307\u6807，\u8fd8\u6709\u4ec0\u4e48\u529e\u6cd5\u907f\u514d\u51c6\u786e\u7387\u9677\u9631？<\/p>\n

A： \u7b2c\u4e00，\u4f7f\u7528\u5206\u5c42\u91c7\u6837（Stratified Sampling）\u5212\u5206\u8bad\u7ec3\/\u6d4b\u8bd5\u96c6，\u786e\u4fdd\u9a8c\u8bc1\u96c6\u4e2d\u7684\u6b63\u6837\u672c\u6bd4\u4f8b\u4e0e\u6574\u4f53\u4e00\u81f4\u3002\u7b2c\u4e8c，\u5728\u8bad\u7ec3\u65f6\u5c31\u7528\u52a0\u6743\u635f\u5931\u6216\u91c7\u6837\u7b56\u7565，\u800c\u4e0d\u662f\u7b49\u6a21\u578b\u8bad\u7ec3\u5b8c\u4e86\u624d\u53d1\u73b0\u6307\u6807\u6709\u95ee\u9898\u3002\u7b2c\u4e09，\u5efa\u7acb\u4e1a\u52a1\u5c42\u9762\u7684\u6536\u76ca\u77e9\u9635\u2014\u2014\u628a\u6df7\u6dc6\u77e9\u9635\u7684\u56db\u4e2a\u8c61\u9650\u6620\u5c04\u5230\u5b9e\u9645\u7684\u4e1a\u52a1\u6210\u672c\u6216\u6536\u76ca\u3002<\/p>\n

Q：\u4e3a\u4ec0\u4e48\u5728\u9ad8\u5ea6\u4e0d\u5e73\u8861\u65f6 AUC-PR \u6bd4 AUC-ROC \u66f4\u53ef\u9760？<\/p>\n

A： ROC \u66f2\u7ebf\u7684\u6a2a\u8f74\u662f FPR = FP \/ (FP + TN)，\u5206\u6bcd\u5305\u542b\u5927\u91cf TN（\u591a\u6570\u7c7b），\u6240\u4ee5 FPR \u5bf9 FP \u7684\u53d8\u5316\u4e0d\u654f\u611f\u3002\u800c PR \u66f2\u7ebf\u7684\u6a2a\u8f74\u662f Recall，\u7eb5\u8f74\u662f Precision，\u76f4\u63a5\u805a\u7126\u4e8e\u6b63\u6837\u672c\u7684\u8bc6\u522b\u8d28\u91cf\u3002\u5f53\u6b63\u6837\u672c\u6781\u5ea6\u7a00\u5c11\u65f6，PR \u66f2\u7ebf\u7684\u4e0b\u964d\u5e45\u5ea6\u8fdc\u6bd4 ROC \u66f2\u7ebf\u660e\u663e，\u66f4\u80fd\u533a\u5206\u6a21\u578b\u597d\u574f\u3002<\/p>\n

\n

2.2 \u6570\u636e\u5c42\u9762\u65b9\u6cd5：\u8fc7\u91c7\u6837\u4e0e\u6b20\u91c7\u6837<\/h3>\n

2.2.1 \u968f\u673a\u6b20\u91c7\u6837（Random Undersampling）<\/h4>\n

\u4ece\u591a\u6570\u7c7b\u4e2d\u968f\u673a\u79fb\u9664\u6837\u672c，\u4f7f\u591a\u6570\u7c7b\u548c\u5c11\u6570\u7c7b\u6570\u91cf\u63a5\u8fd1\u3002<\/p>\n

from<\/span> imblearn.<\/span>under_sampling import<\/span> RandomUnderSampler<\/p>\n

rus =<\/span> RandomUnderSampler(<\/span>random_state=<\/span>42<\/span>,<\/span> sampling_strategy=<\/span>'auto'<\/span>)<\/span>
\nX_res,<\/span> y_res =<\/span> rus.<\/span>fit_resample(<\/span>X,<\/span> y)<\/span>
\nprint<\/span>(<\/span>f'\u91c7\u6837\u524d: <\/span>{<\/span>Counter(<\/span>y)<\/span>}<\/span><\/span>'<\/span><\/span>)<\/span>
\nprint<\/span>(<\/span>f'\u91c7\u6837\u540e: <\/span>{<\/span>Counter(<\/span>y_res)<\/span>}<\/span><\/span>'<\/span><\/span>)<\/span><\/p>\n

\u4f18\u70b9：\u7b80\u5355\u76f4\u63a5\u3001\u51cf\u5c11\u8bad\u7ec3\u6570\u636e\u91cf\u3001\u52a0\u5feb\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n

\u7f3a\u70b9：\u53ef\u80fd\u4e22\u5f03\u6709\u4ef7\u503c\u7684\u591a\u6570\u7c7b\u6837\u672c，\u5bfc\u81f4\u4fe1\u606f\u635f\u5931\u3002\u5f53\u4e0d\u5e73\u8861\u7387\u6781\u9ad8（\u5982 1:10,000）\u65f6，\u6b20\u91c7\u6837\u540e\u7684\u6570\u636e\u91cf\u53ef\u80fd\u592a\u5c11，\u6a21\u578b\u6b20\u62df\u5408\u3002<\/p>\n

2.2.2 \u968f\u673a\u8fc7\u91c7\u6837（Random Oversampling）<\/h4>\n

\u590d\u5236\u5c11\u6570\u7c7b\u6837\u672c，\u4f7f\u5c11\u6570\u7c7b\u548c\u591a\u6570\u7c7b\u6570\u91cf\u63a5\u8fd1\u3002<\/p>\n

from<\/span> imblearn.<\/span>over_sampling import<\/span> RandomOverSampler<\/p>\n

ros =<\/span> RandomOverSampler(<\/span>random_state=<\/span>42<\/span>)<\/span>
\nX_res,<\/span> y_res =<\/span> ros.<\/span>fit_resample(<\/span>X,<\/span> y)<\/span>
\nprint<\/span>(<\/span>f'\u91c7\u6837\u524d: <\/span>{<\/span>Counter(<\/span>y)<\/span>}<\/span><\/span>'<\/span><\/span>)<\/span>
\nprint<\/span>(<\/span>f'\u91c7\u6837\u540e: <\/span>{<\/span>Counter(<\/span>y_res)<\/span>}<\/span><\/span>'<\/span><\/span>)<\/span><\/p>\n

\u4f18\u70b9：\u4e0d\u4e22\u5931\u4fe1\u606f\u3001\u5b9e\u73b0\u7b80\u5355\u3002<\/p>\n

\u7f3a\u70b9：\u53ea\u662f\u7b80\u5355\u590d\u5236\u6837\u672c，\u6a21\u578b\u5bb9\u6613\u8fc7\u62df\u5408\u2014\u2014\u540c\u4e00\u4e2a\u6837\u672c\u89c1\u8fc7\u591a\u6b21，\u51b3\u7b56\u8fb9\u754c\u4f1a\u8fc7\u4e8e\u7d27\u8d34\u8fd9\u4e9b\u590d\u5236\u6837\u672c\u3002<\/p>\n

2.2.3 \u8fc7\u91c7\u6837 vs \u6b20\u91c7\u6837\u5bf9\u6bd4<\/h4>\n\n\u7ef4\u5ea6\u8fc7\u91c7\u6837\u6b20\u91c7\u6837<\/tr>\n\n\n\n\n\n\n\n
\u6570\u636e\u91cf\u53d8\u5316<\/td>\n\u589e\u591a<\/td>\n\u51cf\u5c11<\/td>\n<\/tr>\n
\u4fe1\u606f\u4fdd\u7559<\/td>\n\u4fdd\u7559\u5168\u90e8\u591a\u6570\u7c7b\u4fe1\u606f<\/td>\n\u53ef\u80fd\u4e22\u5f03\u591a\u6570\u7c7b\u4fe1\u606f<\/td>\n<\/tr>\n
\u8fc7\u62df\u5408\u98ce\u9669<\/td>\n\u9ad8（\u590d\u5236\u6837\u672c\u5bfc\u81f4）<\/td>\n\u4f4e<\/td>\n<\/tr>\n
\u6b20\u62df\u5408\u98ce\u9669<\/td>\n\u4f4e<\/td>\n\u9ad8（\u6570\u636e\u91cf\u592a\u5c11）<\/td>\n<\/tr>\n
\u8bad\u7ec3\u65f6\u95f4<\/td>\n\u589e\u52a0<\/td>\n\u51cf\u5c11<\/td>\n<\/tr>\n
\u9002\u7528\u573a\u666f<\/td>\n\u6570\u636e\u603b\u91cf\u5c0f\u3001\u591a\u6570\u7c7b\u53ef\u9760<\/td>\n\u6570\u636e\u603b\u91cf\u5927\u3001\u591a\u6570\u7c7b\u6709\u5197\u4f59<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n

2.3 SMOTE \u53ca\u5176\u53d8\u4f53\u5bb6\u65cf<\/h3>\n

2.3.1 SMOTE \u6838\u5fc3\u539f\u7406<\/h4>\n

SMOTE（Synthetic Minority Over-sampling Technique）\u4e0d\u518d\u662f\u7b80\u5355\u590d\u5236，\u800c\u662f\u5728\u5c11\u6570\u7c7b\u6837\u672c\u4e4b\u95f4\u63d2\u503c\u751f\u6210\u5168\u65b0\u6837\u672c，\u8fd9\u662f\u5b83\u548c\u968f\u673a\u8fc7\u91c7\u6837\u7684\u672c\u8d28\u533a\u522b\u3002<\/p>\n

\u751f\u6210\u6b65\u9aa4：<\/p>\n

  • \u5bf9\u6bcf\u4e2a\u5c11\u6570\u7c7b\u6837\u672c x_i，\u5728\u7279\u5f81\u7a7a\u95f4\u4e2d\u627e\u51fa k \u4e2a\u6700\u8fd1\u90bb（\u540c\u4e3a\u5c11\u6570\u7c7b）<\/li>\n
  • \u4ece k \u4e2a\u8fd1\u90bb\u4e2d\u968f\u673a\u9009\u62e9\u4e00\u4e2a x_nn<\/li>\n
  • \u5728 x_i \u4e0e x_nn \u7684\u8fde\u7ebf\u4e0a\u968f\u673a\u751f\u6210\u4e00\u4e2a\u65b0\u6837\u672c：
    \nxnew=xi+\u03bb\u22c5(xnn\u2212xi),\u03bb\u2208[0,1] x_{new} = x_i + \\\\lambda \\\\cdot (x_{nn} – x_i), \\\\quad \\\\lambda \\\\in [0, 1] <\/span><\/span>x<\/span><\/span>n<\/span>e<\/span>w<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>x<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span><\/span><\/span>\u03bb<\/span><\/span>\u22c5<\/span><\/span><\/span><\/span>(<\/span>x<\/span><\/span>nn<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2212<\/span><\/span><\/span><\/span>x<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>,<\/span><\/span><\/span>\u03bb<\/span><\/span>\u2208<\/span><\/span><\/span><\/span>[<\/span>0<\/span>,<\/span><\/span>1<\/span>]<\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n

    import<\/span> numpy as<\/span> np
    \nfrom<\/span> sklearn.<\/span>neighbors import<\/span> NearestNeighbors<\/p>\n

    def<\/span> smote_generate<\/span>(<\/span>X_minority,<\/span> N,<\/span> k=<\/span>5<\/span>)<\/span>:<\/span>
    \n """\u624b\u52a8\u5b9e\u73b0 SMOTE \u6838\u5fc3\u903b\u8f91（\u7b80\u5316\u7248）"""<\/span>
    \n n_samples,<\/span> n_features =<\/span> X_minority.<\/span>shape
    \n nbrs =<\/span> NearestNeighbors(<\/span>n_neighbors=<\/span>k)<\/span>.<\/span>fit(<\/span>X_minority)<\/span>
    \n synthetic =<\/span> [<\/span>]<\/span><\/p>\n

    for<\/span> _ in<\/span> range<\/span>(<\/span>N)<\/span>:<\/span>
    \n idx =<\/span> np.<\/span>random.<\/span>randint(<\/span>0<\/span>,<\/span> n_samples)<\/span>
    \n x_i =<\/span> X_minority[<\/span>idx]<\/span>
    \n distances,<\/span> indices =<\/span> nbrs.<\/span>kneighbors(<\/span>[<\/span>x_i]<\/span>)<\/span>
    \n nn_idx =<\/span> np.<\/span>random.<\/span>choice(<\/span>indices[<\/span>0<\/span>]<\/span>)<\/span>
    \n x_nn =<\/span> X_minority[<\/span>nn_idx]<\/span>
    \n lam =<\/span> np.<\/span>random.<\/span>random(<\/span>)<\/span>
    \n x_new =<\/span> x_i +<\/span> lam *<\/span> (<\/span>x_nn –<\/span> x_i)<\/span>
    \n synthetic.<\/span>append(<\/span>x_new)<\/span><\/p>\n

    return<\/span> np.<\/span>array(<\/span>synthetic)<\/span><\/p>\n

    # \u5b9e\u9645\u4f7f\u7528\u63a8\u8350\u76f4\u63a5\u8c03 imblearn<\/span>
    \nfrom<\/span> imblearn.<\/span>over_sampling import<\/span> SMOTE<\/p>\n

    smote =<\/span> SMOTE(<\/span>random_state=<\/span>42<\/span>,<\/span> k_neighbors=<\/span>5<\/span>)<\/span>
    \nX_resampled,<\/span> y_resampled =<\/span> smote.<\/span>fit_resample(<\/span>X,<\/span> y)<\/span><\/p>\n

    SMOTE \u7684\u6839\u672c\u7f3a\u9677：<\/p>\n

      \n
    • \u53ea\u5728\u5c11\u6570\u7c7b\u5185\u90e8\u63d2\u503c，\u6ca1\u6709\u8003\u8651\u591a\u6570\u7c7b\u7684\u5206\u5e03，\u53ef\u80fd\u751f\u6210\u4e0e\u591a\u6570\u7c7b\u91cd\u53e0\u7684\u6837\u672c<\/li>\n
    • \u5728\u9ad8\u7ef4\u7a7a\u95f4\u4e2d，\u6b27\u6c0f\u8ddd\u79bb\u4e0d\u518d\u53ef\u9760，\u8fd1\u90bb\u9009\u62e9\u53ef\u80fd\u6ca1\u6709\u610f\u4e49<\/li>\n
    • \u5f53\u4e0d\u5e73\u8861\u7387\u6781\u9ad8\u65f6，\u5c11\u6570\u7c7b\u6837\u672c\u8fc7\u4e8e\u7a00\u758f，\u8fd1\u90bb\u53ef\u80fd\u5e76\u4e0d\u662f\u771f\u6b63\u7684"\u540c\u7c7b"<\/li>\n<\/ul>\n

      2.3.2 SMOTE \u53d8\u4f53\u5168\u5bb6\u798f<\/h4>\n\n\u53d8\u4f53\u6838\u5fc3\u601d\u60f3\u89e3\u51b3\u7684\u95ee\u9898\u9002\u7528\u573a\u666f<\/tr>\n\n\n\n\n\n\n
      Borderline-SMOTE<\/td>\n\u53ea\u5728\u51b3\u7b56\u8fb9\u754c\u9644\u8fd1\u7684\u5c11\u6570\u7c7b\u6837\u672c\u4e0a\u751f\u6210<\/td>\n\u51cf\u5c11\u8fdc\u79bb\u8fb9\u754c\u7684\u566a\u58f0\u6837\u672c<\/td>\n\u8fb9\u754c\u6e05\u6670\u7684\u5206\u7c7b\u95ee\u9898<\/td>\n<\/tr>\n
      ADASYN<\/td>\n\u6839\u636e\u5bc6\u5ea6\u5206\u5e03\u81ea\u9002\u5e94\u751f\u6210\u6570\u91cf\u2014\u2014\u5bc6\u5ea6\u8d8a\u4f4e\u751f\u6210\u8d8a\u591a<\/td>\n\u8ba9\u751f\u6210\u96c6\u4e2d\u5728"\u56f0\u96be\u533a\u57df"<\/td>\n\u975e\u5747\u5300\u5206\u5e03\u7684\u5c11\u6570\u7c7b<\/td>\n<\/tr>\n
      SVMSMOTE<\/td>\n\u7528 SVM \u627e\u51fa\u652f\u6301\u5411\u91cf，\u5728\u652f\u6301\u5411\u91cf\u9644\u8fd1\u751f\u6210<\/td>\n\u805a\u7126\u4e8e\u771f\u6b63\u5f71\u54cd\u51b3\u7b56\u9762\u7684\u6837\u672c<\/td>\n\u9ad8\u7ef4\u7a00\u758f\u6570\u636e<\/td>\n<\/tr>\n
      KMeansSMOTE<\/td>\n\u5148\u7528 KMeans \u805a\u7c7b，\u5728\u6bcf\u4e2a\u7c07\u5185\u72ec\u7acb\u5e94\u7528 SMOTE<\/td>\n\u89e3\u51b3\u591a\u6a21\u6001\u5206\u5e03\u7684\u5c11\u6570\u7c7b<\/td>\n\u5c11\u6570\u7c7b\u6709\u591a\u4e2a\u5b50\u7c07<\/td>\n<\/tr>\n
      SMOTE-ENN<\/td>\nSMOTE \u540e\u7528 Edited Nearest Neighbors \u6e05\u6d17\u566a\u58f0<\/td>\n\u53bb\u9664\u751f\u6210\u540e\u4ecd\u7136\u5206\u7c7b\u9519\u8bef\u7684\u6837\u672c<\/td>\n\u751f\u6210\u6837\u672c\u8d28\u91cf\u5dee\u65f6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      from<\/span> imblearn.<\/span>over_sampling import<\/span> BorderlineSMOTE,<\/span> ADASYN,<\/span> SVMSMOTE,<\/span> KMeansSMOTE
      \nfrom<\/span> imblearn.<\/span>combine import<\/span> SMOTEENN<\/p>\n

      # Borderline-SMOTE: \u5173\u6ce8\u8fb9\u754c\u6837\u672c<\/span>
      \nborderline_smote =<\/span> BorderlineSMOTE(<\/span>kind=<\/span>'borderline-1'<\/span>,<\/span> random_state=<\/span>42<\/span>)<\/span><\/p>\n

      # ADASYN: \u81ea\u9002\u5e94\u5408\u6210\u91c7\u6837<\/span>
      \nadasyn =<\/span> ADASYN(<\/span>random_state=<\/span>42<\/span>,<\/span> n_neighbors=<\/span>5<\/span>)<\/span><\/p>\n

      # SMOTE-ENN: \u7ed3\u5408\u6b20\u91c7\u6837\u6e05\u6d17<\/span>
      \nsmote_enn =<\/span> SMOTEENN(<\/span>random_state=<\/span>42<\/span>)<\/span><\/p>\n

      # \u5b9e\u9645\u4f7f\u7528\u65f6\u6ce8\u610f: SMOTE \u5047\u8bbe\u7279\u5f81\u662f\u8fde\u7eed\u6570\u503c\u578b<\/span>
      \nfrom<\/span> imblearn.<\/span>over_sampling import<\/span> SMOTENC
      \nsmote_nc =<\/span> SMOTENC(<\/span>categorical_features=<\/span>[<\/span>1<\/span>,<\/span> 3<\/span>]<\/span>,<\/span> random_state=<\/span>42<\/span>)<\/span><\/p>\n

      \u9762\u8bd5\u5b98\u8ffd\u95ee\u94fe<\/h4>\n

      Q：SMOTE \u751f\u6210\u7684\u6837\u672c\u53ef\u80fd\u4e0d\u771f\u5b9e，\u600e\u4e48\u5904\u7406？<\/p>\n

      A： \u5bf9\u4e8e\u56fe\u50cf（\u539f\u59cb\u50cf\u7d20\u7a7a\u95f4），\u76f8\u90bb\u6837\u672c\u63d2\u503c\u4f1a\u4ea7\u751f"\u9b3c\u5f71"；\u4f46\u5bf9\u4e8e\u6570\u503c\u578b\u8868\u683c\u6570\u636e，\u63d2\u503c\u901a\u5e38\u5408\u7406\u3002\u5e94\u5bf9\u65b9\u6cd5\u6709：（1）\u4f7f\u7528 SMOTE-ENN \u6216 SMOTE-Tomek \u5728\u751f\u6210\u540e\u6e05\u6d17；（2）\u5728\u751f\u6210\u540e\u505a\u6570\u636e\u9a8c\u8bc1，\u5254\u9664\u4e0e\u591a\u6570\u7c7b\u9ad8\u5ea6\u91cd\u53e0\u7684\u6837\u672c；（3）\u6539\u7528 GAN \u6216 VAE \u751f\u6210\u66f4\u771f\u5b9e\u7684\u6837\u672c\u3002<\/p>\n

      Q：SMOTE \u5bf9\u9ad8\u7ef4\u6570\u636e\u6548\u679c\u5982\u4f55？\u4e3a\u4ec0\u4e48？<\/p>\n

      A： \u9ad8\u7ef4\u7a7a\u95f4\u4e0b\u6548\u679c\u5f80\u5f80\u4e0d\u597d\u3002\u4e3b\u8981\u539f\u56e0\u6709\u4e24\u4e2a：（1）\u7ef4\u5ea6\u8bc5\u5492\u2014\u2014\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u6837\u672c\u6781\u5ea6\u7a00\u758f，\u6b27\u6c0f\u8ddd\u79bb\u8d8b\u4e8e\u4e00\u81f4，\u8fd1\u90bb\u9009\u62e9\u5931\u6548；（2）\u566a\u58f0\u653e\u5927\u2014\u2014\u5728\u9ad8\u7ef4\u7a7a\u95f4\u4e2d\u63d2\u503c\u4f1a\u5f15\u5165\u5927\u91cf\u865a\u5047\u7279\u5f81\u7ec4\u5408\u3002\u89e3\u51b3\u529e\u6cd5：\u5148\u964d\u7ef4（PCA\u3001Autoencoder）\u518d\u505a SMOTE，\u6216\u4f7f\u7528\u57fa\u4e8e GAN \u7684\u65b9\u6cd5\u3002<\/p>\n

      Q：SMOTE \u80fd\u7528\u4e8e NLP \u6216\u65f6\u95f4\u5e8f\u5217\u6570\u636e\u5417？<\/p>\n

      A： \u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528\u3002NLP \u6570\u636e\u7684\u7279\u5f81\u7a7a\u95f4\u662f\u975e\u8fde\u7eed\u3001\u79bb\u6563\u7684，\u63d2\u503c\u5728\u8bed\u4e49\u4e0a\u6ca1\u6709\u610f\u4e49\u3002\u65f6\u95f4\u5e8f\u5217\u7684\u6570\u636e\u5b58\u5728\u65f6\u5e8f\u4f9d\u8d56，\u72ec\u7acb\u63d2\u503c\u4f1a\u7834\u574f\u65f6\u95f4\u7ed3\u6784\u3002NLP \u53ef\u4ee5\u5c1d\u8bd5\u5728 Embedding \u7a7a\u95f4\u505a\u63d2\u503c，\u65f6\u95f4\u5e8f\u5217\u53ef\u4ee5\u5c1d\u8bd5\u5b50\u5e8f\u5217\u7ea7\u522b\u7684 DTW \u63d2\u503c\u6216\u751f\u6210\u5f0f\u65b9\u6cd5\u3002<\/p>\n

      \n

      2.4 Focal Loss \u2014\u2014 \u4ece\u635f\u5931\u51fd\u6570\u5c42\u9762\u89e3\u51b3\u4e0d\u5e73\u8861<\/h3>\n

      2.4.1 \u6807\u51c6\u4ea4\u53c9\u71b5\u7684\u5c40\u9650\u6027<\/h4>\n

      \u4e8c\u5206\u7c7b\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570：<\/p>\n

      CE(p,y)={\u2212log\u2061(p)if\u00a0y=1\u2212log\u2061(1\u2212p)if\u00a0y=0 \\\\text{CE}(p, y) = \\\\begin{cases} -\\\\log(p) & \\\\text{if } y = 1 \\\\\\\\ -\\\\log(1 – p) & \\\\text{if } y = 0 \\\\end{cases} <\/span><\/span>CE<\/span><\/span>(<\/span>p<\/span>,<\/span><\/span>y<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>{<\/span><\/span><\/span>\u2212<\/span><\/span>log<\/span><\/span>(<\/span>p<\/span>)<\/span><\/span><\/span><\/span>\u2212<\/span><\/span>log<\/span><\/span>(<\/span>1<\/span><\/span>\u2212<\/span><\/span>p<\/span>)<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>if\u00a0<\/span><\/span>y<\/span><\/span>=<\/span><\/span>1<\/span><\/span><\/span><\/span>if\u00a0<\/span><\/span>y<\/span><\/span>=<\/span><\/span>0<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

      \u66f4\u7b80\u6d01\u7684\u8868\u8fbe（\u4ee4 p_t = p * y + (1-p) * (1-y)）：<\/p>\n

      CE(p,y)=\u2212log\u2061(pt) \\\\text{CE}(p, y) = -\\\\log(p_t) <\/span><\/span>CE<\/span><\/span>(<\/span>p<\/span>,<\/span><\/span>y<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>\u2212<\/span><\/span>log<\/span><\/span>(<\/span>p<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

      \u5bf9\u4e8e\u591a\u6570\u7c7b（\u8d1f\u6837\u672c），\u6a21\u578b\u5f88\u5bb9\u6613\u5c06\u5176\u9884\u6d4b\u4e3a\u4e00\u4e2a\u9ad8\u7f6e\u4fe1\u5ea6\u7684\u8d1f\u6837\u672c（\u5373 p_t \u63a5\u8fd1 1），\u4f46\u5373\u4f7f\u8fd9\u6837，\u6bcf\u4e2a\u6837\u672c\u7684\u635f\u5931\u4ecd\u7136\u4e3a -log(0.99) \u7ea6\u7b49\u4e8e 0.01\u3002\u5f53\u591a\u6570\u7c7b\u6837\u672c\u6570\u91cf\u5de8\u5927\u65f6，\u8fd9\u4e9b\u5fae\u5c0f\u635f\u5931\u7684\u603b\u548c\u4f1a\u538b\u5012\u5c11\u6570\u7c7b\u6837\u672c\u7684\u8d21\u732e\u3002<\/p>\n

      2.4.2 Focal Loss \u7684\u6838\u5fc3\u601d\u60f3<\/h4>\n

      Focal Loss \u7531 Lin et al. (2017) \u5728\u76ee\u6807\u68c0\u6d4b\u4efb\u52a1\u4e2d\u63d0\u51fa，\u5176\u6838\u5fc3\u601d\u60f3\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e00\u53e5\u8bdd：\u8ba9\u6a21\u578b"\u805a\u7126"\u4e8e\u96be\u5206\u7c7b\u7684\u6837\u672c，\u964d\u4f4e\u6613\u5206\u7c7b\u6837\u672c\u7684\u6743\u91cd\u3002<\/p>\n

      FL(pt)=\u2212\u03b1t(1\u2212pt)\u03b3log\u2061(pt) \\\\text{FL}(p_t) = -\\\\alpha_t (1 – p_t)^\\\\gamma \\\\log(p_t) <\/span><\/span>FL<\/span><\/span>(<\/span>p<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>\u2212<\/span>\u03b1<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>1<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>p<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>\u03b3<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>log<\/span><\/span>(<\/span>p<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

      \u4e24\u4e2a\u5173\u952e\u8bbe\u8ba1：<\/p>\n

    • \u8c03\u8282\u56e0\u5b50 (1 – p_t)^gamma：\u5f53\u6837\u672c\u88ab\u6b63\u786e\u5206\u7c7b\u4e14\u7f6e\u4fe1\u5ea6\u9ad8\u65f6（p_t -> 1），\u56e0\u5b50\u8d8b\u8fd1\u4e8e 0，\u635f\u5931\u88ab\u5927\u5e45\u964d\u4f4e；\u5f53\u6837\u672c\u88ab\u8bef\u5206\u7c7b\u65f6（p_t -> 0），\u56e0\u5b50\u8d8b\u8fd1\u4e8e 1，\u635f\u5931\u51e0\u4e4e\u4e0d\u53d7\u5f71\u54cd\u3002<\/li>\n
    • \u6743\u91cd\u56e0\u5b50 alpha_t：\u7528\u4e8e\u8c03\u6574\u6b63\u8d1f\u6837\u672c\u7684\u6743\u91cd\u6bd4\u4f8b，\u63a7\u5236\u7c7b\u522b\u5e73\u8861\u3002<\/li>\n

      \u4e0d\u540c gamma \u503c\u7684\u6548\u679c：<\/p>\n\ngamma\u6613\u5206\u6837\u672c\u7684\u635f\u5931\u8870\u51cf\u5178\u578b\u573a\u666f<\/tr>\n\n\n\n\n\n\n
      0<\/td>\n\u4e0d\u8870\u51cf（\u7b49\u4ef7\u4e8e CE + alpha）<\/td>\n\u57fa\u7ebf<\/td>\n<\/tr>\n
      0.5<\/td>\n\u6e29\u548c\u8870\u51cf<\/td>\n\u8f7b\u5ea6\u4e0d\u5e73\u8861<\/td>\n<\/tr>\n
      1.0<\/td>\n\u4e2d\u7b49\u8870\u51cf<\/td>\n\u4e2d\u5ea6\u4e0d\u5e73\u8861<\/td>\n<\/tr>\n
      2.0<\/td>\n\u5f3a\u70c8\u8870\u51cf<\/td>\n\u4e25\u91cd\u4e0d\u5e73\u8861（\u9ed8\u8ba4\u503c）<\/td>\n<\/tr>\n
      5.0<\/td>\n\u6781\u7aef\u8870\u51cf<\/td>\n\u6781\u7a00\u758f\u7684\u6b63\u6837\u672c<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      2.4.3 Focal Loss \u7684\u68af\u5ea6\u5206\u6790（\u9762\u8bd5\u91cd\u70b9）<\/h4>\n

      \u4ece\u68af\u5ea6\u89d2\u5ea6\u53ef\u4ee5\u66f4\u76f4\u89c2\u5730\u7406\u89e3 Focal Loss \u4e3a\u4ec0\u4e48\u6709\u6548\u3002\u4ee5 y=1 \u4e3a\u4f8b：<\/p>\n

      \u6807\u51c6\u4ea4\u53c9\u71b5\u7684\u68af\u5ea6：
      \n\u2202CE\u2202p=\u22121p \\\\frac{\\\\partial \\\\text{CE}}{\\\\partial p} = -\\\\frac{1}{p} <\/span><\/span><\/span><\/span>\u2202<\/span>p<\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2202<\/span>CE<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u2212<\/span><\/span><\/span>p<\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

      Focal Loss \u7684\u68af\u5ea6：
      \n\u2202FL\u2202p=\u2212\u03b1t(1\u2212p)\u03b3\u22121[\u03b3plog\u2061(p)+(1\u2212p)] \\\\frac{\\\\partial \\\\text{FL}}{\\\\partial p} = -\\\\alpha_t (1 – p)^{\\\\gamma-1} \\\\left[ \\\\gamma p \\\\log(p) + (1-p) \\\\right] <\/span><\/span><\/span><\/span>\u2202<\/span>p<\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2202<\/span>FL<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u2212<\/span>\u03b1<\/span><\/span>t<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>1<\/span><\/span>\u2212<\/span><\/span><\/span><\/span>p<\/span>)<\/span><\/span>\u03b3<\/span>\u2212<\/span>1<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>[<\/span>\u03b3<\/span>p<\/span><\/span>log<\/span><\/span>(<\/span>p<\/span>)<\/span><\/span>+<\/span><\/span>(<\/span>1<\/span><\/span>\u2212<\/span><\/span>p<\/span>)<\/span>]<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

      \u5f53 p -> 1（\u6613\u5206\u7c7b\u6837\u672c）\u65f6，(1-p)^{gamma-1} \u8d8b\u8fd1\u4e8e 0，\u68af\u5ea6\u88ab\u62c9\u4f4e\u2014\u2014\u6a21\u578b\u4e0d\u518d\u82b1\u7cbe\u529b\u5728\u5df2\u7ecf\u5b66\u4f1a\u7684\u6837\u672c\u4e0a\u3002\u5f53 p -> 0（\u96be\u5206\u7c7b\u6837\u672c）\u65f6，(1-p)^{gamma-1} -> 1，\u68af\u5ea6\u4fdd\u6301\u8f83\u5927\u2014\u2014\u6a21\u578b\u6301\u7eed\u5173\u6ce8\u8fd9\u4e9b\u56f0\u96be\u6837\u672c\u3002<\/p>\n

      \u8fd9\u5c31\u662f Focal Loss \u7684\u7cbe\u9ad3：\u5b83\u4e0d\u662f\u7b80\u5355\u5730\u7ed9\u5c11\u6570\u7c7b\u52a0\u6743\u91cd，\u800c\u662f\u52a8\u6001\u5730\u6839\u636e\u6bcf\u4e2a\u6837\u672c\u7684\u5b66\u4e60\u96be\u5ea6\u6765\u5206\u914d\u6743\u91cd\u3002<\/p>\n

      2.4.4 PyTorch \u5b9e\u73b0<\/h4>\n

      import<\/span> torch
      \nimport<\/span> torch.<\/span>nn as<\/span> nn
      \nimport<\/span> torch.<\/span>nn.<\/span>functional as<\/span> F<\/p>\n

      class<\/span> FocalLoss<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
      \n """
      \n Focal Loss \u7684 PyTorch \u5b9e\u73b0
      \n Args:
      \n alpha: \u7c7b\u522b\u6743\u91cd\u56e0\u5b50 (float)，\u7528\u4e8e\u63a7\u5236\u6b63\u8d1f\u6837\u672c\u6bd4\u4f8b
      \n gamma: \u805a\u7126\u53c2\u6570，gamma >= 0，\u9ed8\u8ba4 2.0
      \n reduction: 'none' | 'mean' | 'sum'
      \n """<\/span>
      \n def<\/span> __init__<\/span>(<\/span>self,<\/span> alpha=<\/span>0.25<\/span>,<\/span> gamma=<\/span>2.0<\/span>,<\/span> reduction=<\/span>'mean'<\/span>)<\/span>:<\/span>
      \n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
      \n self.<\/span>alpha =<\/span> alpha
      \n self.<\/span>gamma =<\/span> gamma
      \n self.<\/span>reduction =<\/span> reduction<\/p>\n

      def<\/span> forward<\/span>(<\/span>self,<\/span> inputs,<\/span> targets)<\/span>:<\/span>
      \n # inputs: raw logits (\u672a\u7ecf\u8fc7 sigmoid)<\/span>
      \n # targets: \u4e8c\u5206\u7c7b\u6807\u7b7e {0, 1}<\/span>
      \n probs =<\/span> torch.<\/span>sigmoid(<\/span>inputs)<\/span>
      \n p_t =<\/span> probs *<\/span> targets +<\/span> (<\/span>1<\/span> –<\/span> probs)<\/span> *<\/span> (<\/span>1<\/span> –<\/span> targets)<\/span><\/p>\n

      # \u8ba1\u7b97\u4ea4\u53c9\u71b5\u635f\u5931: -log(p_t)<\/span>
      \n ce_loss =<\/span> F.<\/span>binary_cross_entropy_with_logits(<\/span>
      \n inputs,<\/span> targets,<\/span> reduction=<\/span>'none'<\/span>
      \n )<\/span><\/p>\n

      # Focal Loss \u6743\u91cd: (1 – p_t) ** gamma<\/span>
      \n focal_weight =<\/span> (<\/span>1<\/span> –<\/span> p_t)<\/span> **<\/span> self.<\/span>gamma<\/p>\n

      # alpha \u5e73\u8861\u6743\u91cd<\/span>
      \n if<\/span> self.<\/span>alpha is<\/span> not<\/span> None<\/span>:<\/span>
      \n alpha_weight =<\/span> targets *<\/span> self.<\/span>alpha +<\/span> (<\/span>1<\/span> –<\/span> targets)<\/span> *<\/span> (<\/span>1<\/span> –<\/span> self.<\/span>alpha)<\/span>
      \n focal_weight =<\/span> focal_weight *<\/span> alpha_weight<\/p>\n

      loss =<\/span> focal_weight *<\/span> ce_loss<\/p>\n

      if<\/span> self.<\/span>reduction ==<\/span> 'mean'<\/span>:<\/span>
      \n return<\/span> loss.<\/span>mean(<\/span>)<\/span>
      \n elif<\/span> self.<\/span>reduction ==<\/span> 'sum'<\/span>:<\/span>
      \n return<\/span> loss.<\/span>sum<\/span>(<\/span>)<\/span>
      \n return<\/span> loss<\/p>\n

      # \u591a\u5206\u7c7b\u7248\u672c<\/span>
      \nclass<\/span> FocalLossMultiClass<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
      \n def<\/span> __init__<\/span>(<\/span>self,<\/span> alpha=<\/span>None<\/span>,<\/span> gamma=<\/span>2.0<\/span>,<\/span> reduction=<\/span>'mean'<\/span>)<\/span>:<\/span>
      \n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
      \n self.<\/span>alpha =<\/span> alpha
      \n self.<\/span>gamma =<\/span> gamma
      \n self.<\/span>reduction =<\/span> reduction<\/p>\n

      def<\/span> forward<\/span>(<\/span>self,<\/span> inputs,<\/span> targets)<\/span>:<\/span>
      \n log_probs =<\/span> F.<\/span>log_softmax(<\/span>inputs,<\/span> dim=<\/span>–<\/span>1<\/span>)<\/span>
      \n probs =<\/span> torch.<\/span>exp(<\/span>log_probs)<\/span>
      \n p_t =<\/span> probs.<\/span>gather(<\/span>1<\/span>,<\/span> targets.<\/span>unsqueeze(<\/span>1<\/span>)<\/span>)<\/span>.<\/span>squeeze(<\/span>1<\/span>)<\/span>
      \n focal_weight =<\/span> (<\/span>1<\/span> –<\/span> p_t)<\/span> **<\/span> self.<\/span>gamma<\/p>\n

      if<\/span> self.<\/span>alpha is<\/span> not<\/span> None<\/span>:<\/span>
      \n alpha_weight =<\/span> self.<\/span>alpha.<\/span>gather(<\/span>0<\/span>,<\/span> targets)<\/span>
      \n focal_weight =<\/span> focal_weight *<\/span> alpha_weight<\/p>\n

      loss =<\/span> focal_weight *<\/span> (<\/span>–<\/span>log_probs.<\/span>gather(<\/span>1<\/span>,<\/span> targets.<\/span>unsqueeze(<\/span>1<\/span>)<\/span>)<\/span>.<\/span>squeeze(<\/span>1<\/span>)<\/span>)<\/span><\/p>\n

      if<\/span> self.<\/span>reduction ==<\/span> 'mean'<\/span>:<\/span>
      \n return<\/span> loss.<\/span>mean(<\/span>)<\/span>
      \n elif<\/span> self.<\/span>reduction ==<\/span> 'sum'<\/span>:<\/span>
      \n return<\/span> loss.<\/span>sum<\/span>(<\/span>)<\/span>
      \n return<\/span> loss<\/p>\n

      \u9762\u8bd5\u5b98\u8ffd\u95ee\u94fe<\/h4>\n

      Q：Focal Loss \u548c\u52a0\u6743\u7684 Cross Entropy \u6709\u4ec0\u4e48\u533a\u522b？<\/p>\n

      A： \u52a0\u6743 CE \u662f\u9759\u6001\u7684\u2014\u2014\u6bcf\u4e2a\u6b63\u6837\u672c\u7684\u6743\u91cd\u7cfb\u6570\u662f\u56fa\u5b9a\u7684（\u4f8b\u5982\u6b63\u6837\u672c\u6743\u91cd=100，\u8d1f\u6837\u672c\u6743\u91cd=1），\u4e0d\u5173\u5fc3\u6837\u672c\u672c\u8eab\u662f\u5426\u5bb9\u6613\u88ab\u5206\u7c7b\u3002Focal Loss \u662f\u52a8\u6001\u7684\u2014\u2014\u6743\u91cd (1-p_t)^gamma \u968f\u8bad\u7ec3\u52a8\u6001\u53d8\u5316：\u5df2\u7ecf\u662f\u9ad8\u7f6e\u4fe1\u5ea6\u7684\u6b63\u6837\u672c\u6743\u91cd\u964d\u4f4e，\u4f4e\u7f6e\u4fe1\u5ea6\u7684\u56f0\u96be\u6837\u672c\u6743\u91cd\u4fdd\u6301\u3002\u6362\u53e5\u8bdd\u8bf4，\u52a0\u6743 CE \u53ea\u5728\u6837\u672c\u7c7b\u522b\u7ef4\u5ea6\u4e0a\u505a\u533a\u522b\u5bf9\u5f85，Focal Loss \u5728\u6837\u672c\u4e2a\u4f53\u96be\u5ea6\u7ef4\u5ea6\u4e0a\u505a\u533a\u522b\u5bf9\u5f85\u3002<\/p>\n

      Q：Focal Loss \u4e2d\u7684 alpha \u548c\u6837\u672c\u6743\u91cd\u6709\u4ec0\u4e48\u5173\u7cfb？\u4e3a\u4ec0\u4e48\u4e0d\u53ea\u7528\u4e00\u4e2a gamma \u5c31\u591f\u4e86？<\/p>\n

      A： gamma \u63a7\u5236\u7684\u662f"\u805a\u7126\u7a0b\u5ea6"，\u89e3\u51b3\u7684\u662f\u96be\u6613\u6837\u672c\u4e0d\u5e73\u8861\u95ee\u9898；alpha \u63a7\u5236\u7684\u662f\u6b63\u8d1f\u6837\u672c\u6570\u91cf\u4e0d\u5e73\u8861\u95ee\u9898\u3002\u4e8c\u8005\u89e3\u51b3\u7684\u662f\u4e0d\u540c\u5c42\u9762\u7684\u4e0d\u5e73\u8861\u3002\u5728\u5b9e\u9645\u4f7f\u7528\u4e2d，\u4e24\u8005\u914d\u5408\u6548\u679c\u6700\u597d：alpha \u5e73\u8861\u6570\u91cf，gamma \u8ba9\u6a21\u578b\u5173\u6ce8\u56f0\u96be\u6837\u672c\u3002\u5982\u679c\u4e0d\u8bbe alpha，\u591a\u6570\u7c7b\u4e2d\u4ecd\u7136\u6709\u5927\u91cf\u4e2d\u7b49\u96be\u5ea6\u7684\u6837\u672c\u8d21\u732e\u663e\u8457\u7684\u635f\u5931\u603b\u548c\u3002<\/p>\n

      Q：Focal Loss \u662f\u5426\u53ef\u4ee5\u63a8\u5e7f\u5230\u56de\u5f52\u95ee\u9898？<\/p>\n

      A： \u53ef\u4ee5\u3002\u57fa\u672c\u601d\u8def\u662f：\u5bf9\u6b8b\u5dee\u5c0f\u7684\u6837\u672c（\u6613\u62df\u5408\u7684\u6837\u672c）\u964d\u4f4e\u6743\u91cd，\u5bf9\u6b8b\u5dee\u5927\u7684\u6837\u672c（\u5f02\u5e38\u503c\u6216\u6781\u5c11\u6570\u6837\u672c）\u4fdd\u6301\u6216\u589e\u52a0\u6743\u91cd\u3002\u4f8b\u5982：FL-L1 = (1 – e^{-|y – y_hat|})^gamma * |y – y_hat|\u3002<\/p>\n

      \n

      2.5 \u4ee3\u4ef7\u654f\u611f\u5b66\u4e60 \u2014\u2014 \u8ba9\u6a21\u578b"\u77e5\u9053"\u72af\u9519\u4ee3\u4ef7\u4e0d\u540c<\/h3>\n

      2.5.1 \u6838\u5fc3\u601d\u60f3<\/h4>\n

      \u4ee3\u4ef7\u654f\u611f\u5b66\u4e60（Cost-Sensitive Learning）\u4e0d\u6539\u53d8\u6570\u636e\u7684\u5206\u5e03，\u800c\u662f\u901a\u8fc7\u4fee\u6539\u5b66\u4e60\u76ee\u6807，\u8ba9\u6a21\u578b\u610f\u8bc6\u5230"\u9519\u5224\u5c11\u6570\u7c7b\u7684\u4ee3\u4ef7\u8fdc\u9ad8\u4e8e\u9519\u5224\u591a\u6570\u7c7b"\u3002<\/p>\n

      \u4ee3\u4ef7\u77e9\u9635：\u5b9a\u4e49\u4e00\u4e2a C x C \u7684\u77e9\u9635（C \u4e3a\u7c7b\u522b\u6570），\u5176\u4e2d c_ij \u8868\u793a\u5c06\u771f\u5b9e\u7c7b\u522b i \u9884\u6d4b\u4e3a\u7c7b\u522b j \u7684\u4ee3\u4ef7\u3002<\/p>\n

      \u5bf9\u4e8e\u4e8c\u5206\u7c7b\u95ee\u9898：<\/p>\n\n\u771f\u5b9e\\\\\u9884\u6d4b\u9884\u6d4b\u4e3a\u6b63\u9884\u6d4b\u4e3a\u8d1f<\/tr>\n\n\n\n
      \u6b63\u7c7b<\/td>\nc_11 = 0<\/td>\nc_10 = FN_cost<\/td>\n<\/tr>\n
      \u8d1f\u7c7b<\/td>\nc_01 = FP_cost<\/td>\nc_00 = 0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      \u76ee\u6807\u662f\u6700\u5c0f\u5316\u671f\u671b\u4ee3\u4ef7：R(model) = E[ c_y_y_hat ]<\/p>\n

      2.5.2 \u5b9e\u73b0\u65b9\u5f0f<\/h4>\n

      \u65b9\u5f0f\u4e00：\u5728\u635f\u5931\u51fd\u6570\u4e2d\u52a0\u6743<\/p>\n

      import<\/span> torch.<\/span>nn as<\/span> nn<\/p>\n

      # \u6b63\u6837\u672c\u6743\u91cd = \u8d1f\u6837\u672c\u6570 \/ \u6b63\u6837\u672c\u6570<\/span>
      \npos_weight =<\/span> torch.<\/span>tensor(<\/span>[<\/span>n_negative \/<\/span> n_positive]<\/span>)<\/span>
      \ncriterion =<\/span> nn.<\/span>BCEWithLogitsLoss(<\/span>pos_weight=<\/span>pos_weight)<\/span><\/p>\n

      \u65b9\u5f0f\u4e8c：\u5728\u6a21\u578b\u8f93\u51fa\u540e\u8c03\u6574\u9608\u503c<\/p>\n

      \u4f20\u7edf\u5206\u7c7b\u5668\u4f7f\u7528 0.5 \u4f5c\u4e3a\u51b3\u7b56\u9608\u503c\u3002\u5bf9\u4e8e\u4e0d\u5e73\u8861\u6570\u636e，\u964d\u4f4e\u9608\u503c\u53ef\u4ee5\u8ba9\u66f4\u591a\u6837\u672c\u88ab\u5206\u5230\u5c11\u6570\u7c7b：<\/p>\n

      from<\/span> sklearn.<\/span>ensemble import<\/span> RandomForestClassifier<\/p>\n

      model =<\/span> RandomForestClassifier(<\/span>)<\/span>
      \nmodel.<\/span>fit(<\/span>X_train,<\/span> y_train)<\/span><\/p>\n

      y_prob =<\/span> model.<\/span>predict_proba(<\/span>X_test)<\/span>[<\/span>:<\/span>,<\/span> 1<\/span>]<\/span><\/p>\n

      # \u5728\u9a8c\u8bc1\u96c6\u4e0a\u641c\u7d22\u6700\u4f18\u9608\u503c<\/span>
      \nfrom<\/span> sklearn.<\/span>metrics import<\/span> f1_score<\/p>\n

      thresholds =<\/span> np.<\/span>linspace(<\/span>0.05<\/span>,<\/span> 0.95<\/span>,<\/span> 100<\/span>)<\/span>
      \nbest_thresh =<\/span> 0.5<\/span>
      \nbest_f1 =<\/span> 0.0<\/span><\/p>\n

      for<\/span> thresh in<\/span> thresholds:<\/span>
      \n y_pred =<\/span> (<\/span>y_prob >=<\/span> thresh)<\/span>.<\/span>astype(<\/span>int<\/span>)<\/span>
      \n f1 =<\/span> f1_score(<\/span>y_val,<\/span> y_pred)<\/span>
      \n if<\/span> f1 ><\/span> best_f1:<\/span>
      \n best_f1 =<\/span> f1
      \n best_thresh =<\/span> thresh<\/p>\n

      print<\/span>(<\/span>f"\u6700\u4f18\u9608\u503c: <\/span>{<\/span>best_thresh:<\/span>.3f<\/span>}<\/span><\/span>, \u6700\u4f18 F1: <\/span>{<\/span>best_f1:<\/span>.4f<\/span>}<\/span><\/span>"<\/span><\/span>)<\/span><\/p>\n

      \u65b9\u5f0f\u4e09：\u5728\u6a21\u578b\u4e2d\u5d4c\u5165\u4ee3\u4ef7\u4fe1\u606f<\/p>\n

      # XGBoost \u4e2d\u8bbe\u7f6e scale_pos_weight<\/span>
      \nimport<\/span> xgboost as<\/span> xgb<\/p>\n

      scale_pos_weight =<\/span> n_negative \/<\/span> n_positive
      \nmodel =<\/span> xgb.<\/span>XGBClassifier(<\/span>
      \n scale_pos_weight=<\/span>scale_pos_weight,<\/span>
      \n eval_metric=<\/span>'auc'<\/span>
      \n)<\/span><\/p>\n

      # LightGBM \u4e2d\u8bbe\u7f6e class_weight<\/span>
      \nimport<\/span> lightgbm as<\/span> lgb<\/p>\n

      model =<\/span> lgb.<\/span>LGBMClassifier(<\/span>
      \n class_weight=<\/span>'balanced'<\/span>,<\/span>
      \n scale_pos_weight=<\/span>scale_pos_weight
      \n)<\/span><\/p>\n

      2.5.3 \u6570\u636e\u91c7\u6837 vs \u4ee3\u4ef7\u654f\u611f vs \u635f\u5931\u51fd\u6570\u65b9\u6cd5\u7684\u5bf9\u6bd4<\/h4>\n\n\u65b9\u6cd5\u539f\u7406\u5c42\u9762\u662f\u5426\u6539\u6570\u636e\u8bad\u7ec3\u5f00\u9500\u53ef\u89e3\u91ca\u6027\u8c03\u53c2\u96be\u5ea6<\/tr>\n\n\n\n\n\n\n\n
      \u8fc7\u91c7\u6837<\/td>\n\u6570\u636e\u5c42\u9762<\/td>\n\u662f<\/td>\n\u7565\u589e<\/td>\n\u9ad8<\/td>\n\u4f4e<\/td>\n<\/tr>\n
      \u6b20\u91c7\u6837<\/td>\n\u6570\u636e\u5c42\u9762<\/td>\n\u662f<\/td>\n\u51cf\u5c11<\/td>\n\u9ad8<\/td>\n\u4f4e<\/td>\n<\/tr>\n
      SMOTE<\/td>\n\u6570\u636e\u5c42\u9762<\/td>\n\u662f<\/td>\n\u589e\u52a0<\/td>\n\u4e2d<\/td>\n\u4e2d<\/td>\n<\/tr>\n
      \u4ee3\u4ef7\u654f\u611f<\/td>\n\u7b97\u6cd5\u5c42\u9762<\/td>\n\u5426<\/td>\n\u51e0\u4e4e\u4e0d\u53d8<\/td>\n\u9ad8<\/td>\n\u4f4e<\/td>\n<\/tr>\n
      Focal Loss<\/td>\n\u635f\u5931\u51fd\u6570<\/td>\n\u5426<\/td>\n\u51e0\u4e4e\u4e0d\u53d8<\/td>\n\u4e2d<\/td>\n\u4e2d<\/td>\n<\/tr>\n
      \u9608\u503c\u8c03\u6574<\/td>\n\u540e\u5904\u7406<\/td>\n\u5426<\/td>\n\u65e0<\/td>\n\u9ad8<\/td>\n\u4e2d<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

      \u9762\u8bd5\u5b98\u8ffd\u95ee\u94fe<\/h4>\n

      Q：\u4ee3\u4ef7\u654f\u611f\u5b66\u4e60\u548c Focal Loss \u6709\u4ec0\u4e48\u672c\u8d28\u533a\u522b？<\/p>\n

      A： \u4ee3\u4ef7\u654f\u611f\u5b66\u4e60\u662f"\u7c7b\u522b\u7ef4\u5ea6"\u7684\u52a0\u6743\u2014\u2014\u540c\u4e00\u7c7b\u522b\u7684\u6240\u6709\u6837\u672c\u5171\u4eab\u4e00\u4e2a\u6743\u91cd，\u6743\u91cd\u7531\u4e1a\u52a1\u7ed9\u51fa\u7684\u4ee3\u4ef7\u77e9\u9635\u51b3\u5b9a\u3002Focal Loss \u662f"\u6837\u672c\u7ef4\u5ea6"\u7684\u52a0\u6743\u2014\u2014\u6bcf\u4e2a\u6837\u672c\u7684\u6743\u91cd\u53d6\u51b3\u4e8e\u5b83\u5f53\u524d\u88ab\u5206\u7c7b\u7684\u96be\u6613\u7a0b\u5ea6，\u968f\u8bad\u7ec3\u52a8\u6001\u53d8\u5316\u3002\u4ece\u6570\u5b66\u4e0a\u8bf4，\u4ee3\u4ef7\u654f\u611f\u5b66\u4e60\u76f8\u5f53\u4e8e\u56fa\u5b9a\u4e86 alpha \u6743\u91cd（gamma=0 \u7684 Focal Loss），\u800c Focal Loss \u662f\u4ee3\u4ef7\u654f\u611f\u7684\u63a8\u5e7f\u3002<\/p>\n

      Q：\u9608\u503c\u79fb\u52a8（Threshold Moving）\u5728\u4ec0\u4e48\u60c5\u51b5\u4e0b\u4f1a\u5931\u6548？<\/p>\n

      A： \u9608\u503c\u79fb\u52a8\u5047\u8bbe\u6a21\u578b\u8f93\u51fa\u7684\u6982\u7387\u662f\u826f\u597d\u6821\u51c6\u7684（well-calibrated）\u3002\u5982\u679c\u6a21\u578b\u8f93\u51fa\u6982\u7387\u6821\u51c6\u4e0d\u826f（\u5982 SVM \u7684 decision function\u3001\u672a\u7ecf\u6e29\u5ea6\u7f29\u653e\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b），\u8c03\u6574\u9608\u503c\u7684\u610f\u4e49\u4e0d\u5927\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\u9700\u8981\u5148\u505a\u6982\u7387\u6821\u51c6（Platt Scaling \u6216 Isotonic Regression），\u6216\u8005\u76f4\u63a5\u4f7f\u7528\u4ee3\u4ef7\u654f\u611f\u5b66\u4e60\u3002<\/p>\n

      \n

      2.6 \u96c6\u6210\u65b9\u6cd5 \u2014\u2014 EasyEnsemble \u4e0e BalanceCascade<\/h3>\n

      2.6.1 \u4e3a\u4ec0\u4e48\u9700\u8981\u96c6\u6210\u65b9\u6cd5？<\/h4>\n

      \u5f53\u4e0d\u5e73\u8861\u7387\u6781\u9ad8（\u5982 1:10,000）\u65f6，\u5355\u6b21\u6b20\u91c7\u6837\u4f1a\u4e22\u5f03\u5927\u91cf\u591a\u6570\u7c7b\u4fe1\u606f，\u800c\u5355\u6b21\u8fc7\u91c7\u6837\u4f1a\u5f15\u5165\u5927\u91cf\u566a\u58f0\u6216\u5bfc\u81f4\u8fc7\u62df\u5408\u3002\u96c6\u6210\u65b9\u6cd5\u901a\u8fc7\u591a\u6b21\u91c7\u6837 + \u591a\u4e2a\u57fa\u5206\u7c7b\u5668\u6765\u7efc\u5408\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n

      2.6.2 EasyEnsemble<\/h4>\n

      EasyEnsemble \u7684\u505a\u6cd5\u975e\u5e38\u76f4\u89c2：<\/p>\n

    • \u4ece\u591a\u6570\u7c7b\u4e2d\u6709\u653e\u56de\u5730\u91c7\u6837\u51fa T \u4e2a\u5b50\u96c6，\u6bcf\u4e2a\u5b50\u96c6\u7684\u6837\u672c\u6570\u91cf\u4e0e\u5c11\u6570\u7c7b\u76f8\u540c<\/li>\n
    • \u7528\u5c11\u6570\u7c7b + \u6bcf\u4e2a\u591a\u6570\u7c7b\u5b50\u96c6\u8bad\u7ec3\u4e00\u4e2a\u57fa\u5206\u7c7b\u5668（\u5171 T \u4e2a）<\/li>\n
    • \u5bf9 T \u4e2a\u5206\u7c7b\u5668\u7684\u9884\u6d4b\u7ed3\u679c\u53d6\u5e73\u5747\u6216\u6295\u7968<\/li>\n

      \u8fd9\u6837，\u6bcf\u4e2a\u57fa\u5206\u7c7b\u5668\u90fd\u5728\u5e73\u8861\u7684\u6570\u636e\u4e0a\u8bad\u7ec3，\u800c\u96c6\u6210\u540e\u7684\u6a21\u578b\u5229\u7528\u4e86\u5168\u90e8\u591a\u6570\u7c7b\u4fe1\u606f\u3002<\/p>\n

      import<\/span> numpy as<\/span> np
      \nfrom<\/span> sklearn.<\/span>tree import<\/span> DecisionTreeClassifier
      \nfrom<\/span> sklearn.<\/span>base import<\/span> BaseEstimator,<\/span> ClassifierMixin
      \nfrom<\/span> sklearn.<\/span>utils import<\/span> resample<\/p>\n

      class<\/span> EasyEnsemble<\/span>(<\/span>BaseEstimator,<\/span> ClassifierMixin)<\/span>:<\/span>
      \n """EasyEnsemble \u624b\u52a8\u5b9e\u73b0"""<\/span>
      \n def<\/span> __init__<\/span>(<\/span>self,<\/span> base_estimator=<\/span>None<\/span>,<\/span> n_estimators=<\/span>10<\/span>)<\/span>:<\/span>
      \n self.<\/span>n_estimators =<\/span> n_estimators
      \n self.<\/span>base_estimator =<\/span> base_estimator or<\/span> DecisionTreeClassifier(<\/span>max_depth=<\/span>3<\/span>)<\/span>
      \n self.<\/span>estimators_ =<\/span> [<\/span>]<\/span><\/p>\n

      def<\/span> fit<\/span>(<\/span>self,<\/span> X,<\/span> y)<\/span>:<\/span>
      \n X_pos =<\/span> X[<\/span>y ==<\/span> 1<\/span>]<\/span>
      \n X_neg =<\/span> X[<\/span>y ==<\/span> 0<\/span>]<\/span>
      \n n_pos =<\/span> len<\/span>(<\/span>X_pos)<\/span><\/p>\n

      for<\/span> _ in<\/span> range<\/span>(<\/span>self.<\/span>n_estimators)<\/span>:<\/span>
      \n X_neg_sample =<\/span> resample(<\/span>X_neg,<\/span> n_samples=<\/span>n_pos,<\/span> random_state=<\/span>None<\/span>)<\/span>
      \n X_bal =<\/span> np.<\/span>vstack(<\/span>[<\/span>X_pos,<\/span> X_neg_sample]<\/span>)<\/span>
      \n y_bal =<\/span> np.<\/span>hstack(<\/span>[<\/span>np.<\/span>ones(<\/span>n_pos)<\/span>,<\/span> np.<\/span>zeros(<\/span>n_pos)<\/span>]<\/span>)<\/span><\/p>\n

      clf =<\/span> self.<\/span>base_estimator.<\/span>__class__(<\/span>**<\/span>self.<\/span>base_estimator.<\/span>get_params(<\/span>)<\/span>)<\/span>
      \n clf.<\/span>fit(<\/span>X_bal,<\/span> y_bal)<\/span>
      \n self.<\/span>estimators_.<\/span>append(<\/span>clf)<\/span>
      \n return<\/span> self<\/p>\n

      def<\/span> predict_proba<\/span>(<\/span>self,<\/span> X)<\/span>:<\/span>
      \n probs =<\/span> np.<\/span>mean(<\/span>[<\/span>clf.<\/span>predict_proba(<\/span>X)<\/span>[<\/span>:<\/span>,<\/span> 1<\/span>]<\/span> for<\/span> clf in<\/span> self.<\/span>estimators_]<\/span>,<\/span> axis=<\/span>0<\/span>)<\/span>
      \n return<\/span> np.<\/span>vstack(<\/span>[<\/span>1<\/span> –<\/span> probs,<\/span> probs]<\/span>)<\/span>.<\/span>T<\/p>\n

      def<\/span> predict<\/span>(<\/span>self,<\/span> X)<\/span>:<\/span>
      \n return<\/span> (<\/span>self.<\/span>predict_proba(<\/span>X)<\/span>[<\/span>:<\/span>,<\/span> 1<\/span>]<\/span> ><\/span> 0.5<\/span>)<\/span>.<\/span>astype(<\/span>int<\/span>)<\/span><\/p>\n

      # \u4f7f\u7528 imbalanced-learn \u4e2d\u7684 EasyEnsemble<\/span>
      \nfrom<\/span> imblearn.<\/span>ensemble import<\/span> EasyEnsembleClassifier<\/p>\n

      eec =<\/span> EasyEnsembleClassifier(<\/span>
      \n n_estimators=<\/span>10<\/span>,<\/span>
      \n base_estimator=<\/span>DecisionTreeClassifier(<\/span>max_depth=<\/span>3<\/span>)<\/span>,<\/span>
      \n sampling_strategy=<\/span>'auto'<\/span>,<\/span>
      \n random_state=<\/span>42<\/span>
      \n)<\/span>
      \neec.<\/span>fit(<\/span>X_train,<\/span> y_train)<\/span><\/p>\n

      2.6.3 BalanceCascade<\/h4>\n

      BalanceCascade \u5728 EasyEnsemble \u7684\u57fa\u7840\u4e0a\u505a\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u6539\u8fdb：\u7ea7\u8054\u6dd8\u6c70\u3002\u6bcf\u8bad\u7ec3\u5b8c\u4e00\u4e2a\u57fa\u5206\u7c7b\u5668\u540e，\u4ece\u591a\u6570\u7c7b\u4e2d\u79fb\u9664\u90a3\u4e9b\u88ab\u5f53\u524d\u5206\u7c7b\u5668\u6b63\u786e\u5206\u7c7b\u7684\u6837\u672c，\u8ba9\u4e0b\u4e00\u4e2a\u5206\u7c7b\u5668\u4e13\u6ce8\u4e8e"\u66f4\u96be"\u7684\u591a\u6570\u7c7b\u6837\u672c\u3002<\/p>\n

      \u6b65\u9aa4 1: \u4ece\u591a\u6570\u7c7b\u4e2d\u91c7\u6837\u4e00\u4e2a\u5927\u5c0f\u4e3a n_pos \u7684\u5b50\u96c6
      \n\u6b65\u9aa4 2: \u7528\u5c11\u6570\u7c7b + \u8be5\u5b50\u96c6\u8bad\u7ec3\u57fa\u5206\u7c7b\u5668
      \n\u6b65\u9aa4 3: \u4ece\u591a\u6570\u7c7b\u4e2d\u79fb\u9664\u8be5\u57fa\u5206\u7c7b\u5668\u80fd\u6b63\u786e\u5206\u7c7b\u7684\u6837\u672c（\u7ea7\u8054）
      \n\u6b65\u9aa4 4: \u91cd\u590d\u6b65\u9aa4 1-3，\u76f4\u5230\u8fbe\u5230 n_estimators \u6216\u591a\u6570\u7c7b\u8017\u5c3d<\/p>\n

      \u8fd9\u6837\u505a\u7684\u76f4\u89c9\u662f：\u5df2\u7ecf\u88ab\u6b63\u786e\u5206\u7c7b\u7684\u591a\u6570\u7c7b\u6837\u672c\u4e0d\u9700\u8981\u518d\u88ab\u540e\u9762\u7684\u5206\u7c7b\u5668\u5173\u6ce8，\u800c"\u96be\u7684"\u591a\u6570\u7c7b\u6837\u672c（\u548c\u5c11\u6570\u7c7b\u6df7\u6dc6\u7684）\u9700\u8981\u66f4\u591a\u5206\u7c7b\u5668\u6765\u8fa8\u6790\u3002<\/p>\n

      from<\/span> imblearn.<\/span>ensemble import<\/span> BalancedBaggingClassifier,<\/span> BalancedRandomForestClassifier<\/p>\n

      # Balanced Bagging: \u6bcf\u4e2a\u57fa\u5b66\u4e60\u5668\u4f7f\u7528\u968f\u673a\u8fc7\u91c7\u6837<\/span>
      \nbbc =<\/span> BalancedBaggingClassifier(<\/span>
      \n base_estimator=<\/span>DecisionTreeClassifier(<\/span>)<\/span>,<\/span>
      \n n_estimators=<\/span>50<\/span>,<\/span>
      \n sampling_strategy=<\/span>'auto'<\/span>,<\/span>
      \n replacement=<\/span>True<\/span>,<\/span>
      \n random_state=<\/span>42<\/span>
      \n)<\/span><\/p>\n

      # Balanced Random Forest: \u6bcf\u4e2a\u6811\u5728\u5e73\u8861\u5b50\u96c6\u4e0a\u8bad\u7ec3<\/span>
      \nbrf =<\/span> BalancedRandomForestClassifier(<\/span>
      \n n_estimators=<\/span>100<\/span>,<\/span>
      \n max_depth=<\/span>10<\/span>,<\/span>
      \n sampling_strategy=<\/span>'auto'<\/span>,<\/span>
      \n random_state=<\/span>42<\/span>
      \n)<\/span><\/p>\n

      \u9762\u8bd5\u5b98\u8ffd\u95ee\u94fe<\/h4>\n

      Q：EasyEnsemble \u548c\u666e\u901a\u7684 Bagging \u6709\u4ec0\u4e48\u533a\u522b？<\/p>\n

      A： \u666e\u901a Bagging \u662f\u5728\u5168\u91cf\u6570\u636e\u4e0a Bootstrap \u91c7\u6837（\u6bcf\u4efd\u6570\u636e\u5206\u5e03\u4e0e\u539f\u59cb\u6570\u636e\u4fdd\u6301\u4e00\u81f4，\u5373\u8fd8\u662f\u4e0d\u5e73\u8861\u7684），\u6bcf\u4e2a\u57fa\u5b66\u4e60\u5668\u9762\u5bf9\u7684\u8fd8\u662f\u4e0d\u5e73\u8861\u6570\u636e\u3002EasyEnsemble \u7684\u6bcf\u4e2a\u5b50\u96c6\u90fd\u662f\u5e73\u8861\u7684（\u5c11\u6570\u7c7b + \u7b49\u91cf\u591a\u6570\u7c7b），\u6240\u4ee5\u6bcf\u4e2a\u57fa\u5b66\u4e60\u5668\u9762\u5bf9\u7684\u662f\u5e73\u8861\u6570\u636e\u2014\u2014\u5b83"\u88ab\u8feb"\u5b66\u4e60\u5c11\u6570\u7c7b\u7684\u6a21\u5f0f\u3002\u6b64\u5916，EasyEnsemble \u4e2d\u591a\u6570\u7c7b\u662f\u6709\u653e\u56de\u91c7\u6837，\u800c\u5c11\u6570\u7c7b\u662f\u4e0d\u53d8\u5730\u590d\u5236\u5230\u6bcf\u4e2a\u5b50\u96c6\u4e2d\u3002<\/p>\n

      Q：\u4e3a\u4ec0\u4e48\u4e0d\u76f4\u63a5\u7528 AdaBoost \u89e3\u51b3\u4e0d\u5e73\u8861？<\/p>\n

      A： \u4f20\u7edf AdaBoost \u4f1a\u7ed9\u5206\u9519\u7684\u6837\u672c\u589e\u52a0\u6743\u91cd\u3002\u5728\u4e0d\u5e73\u8861\u6570\u636e\u4e0a，\u591a\u6570\u7c7b\u7684\u5206\u9519\u6837\u672c\u6570\u91cf\u8fdc\u591a\u4e8e\u5c11\u6570\u7c7b，\u6240\u4ee5\u6743\u91cd\u4f1a\u4e0d\u6210\u6bd4\u4f8b\u5730\u96c6\u4e2d\u5728\u591a\u6570\u7c7b\u4e0a，\u53cd\u800c\u8ba9\u6a21\u578b\u66f4\u504f\u5411\u591a\u6570\u7c7b\u3002\u6539\u8fdb\u65b9\u6cd5\u6709 RUSBoost（\u5148\u6b20\u91c7\u6837\u518d\u505a Boosting）\u548c SMOTEBoost（\u5148 SMOTE \u518d\u505a Boosting），\u5b83\u4eec\u628a\u91c7\u6837\u548c Boosting \u7ec4\u5408\u8d77\u6765\u907f\u514d\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n

      \n

      2.7 \u5f02\u5e38\u68c0\u6d4b\u65b9\u6cd5 \u2014\u2014 \u628a\u4e0d\u5e73\u8861\u95ee\u9898\u8f6c\u5316\u4e3a\u5f02\u5e38\u53d1\u73b0<\/h3>\n

      2.7.1 \u5f53\u4e0d\u5e73\u8861\u7387\u6781\u9ad8\u65f6<\/h4>\n

      \u5f53\u6b63\u8d1f\u6837\u672c\u6bd4\u4f8b\u8fbe\u5230 1:100,000 \u751a\u81f3\u66f4\u4f4e\u65f6，\u4f20\u7edf\u7684\u91c7\u6837 + \u5206\u7c7b\u65b9\u6cd5\u57fa\u672c\u5931\u6548：<\/p>\n

        \n
      • \u6b20\u91c7\u6837\u540e\u6570\u636e\u91cf\u592a\u5c11，\u6a21\u578b\u65e0\u6cd5\u5b66\u5230\u6709\u6548\u6a21\u5f0f<\/li>\n
      • \u8fc7\u91c7\u6837\u751f\u6210\u7684\u6837\u672c\u6570\u91cf\u8fdc\u8d85\u771f\u5b9e\u6b63\u6837\u672c，\u5f15\u5165\u5927\u91cf\u566a\u58f0<\/li>\n
      • SMOTE \u5728\u6781\u7a00\u758f\u7684\u6b63\u6837\u672c\u7a7a\u95f4\u4e2d\u63d2\u503c，\u751f\u6210\u7684\u6837\u672c\u8d28\u91cf\u6781\u4f4e<\/li>\n<\/ul>\n

        \u6b64\u65f6，\u53ef\u4ee5\u5c06\u95ee\u9898\u91cd\u65b0\u5b9a\u4e49\u4e3a\u5f02\u5e38\u68c0\u6d4b（Anomaly Detection）\u2014\u2014\u628a\u5360\u6781\u5c11\u6570\u7684\u6b63\u6837\u672c\u770b\u4f5c"\u5f02\u5e38"，\u628a\u591a\u6570\u7c7b\u770b\u4f5c"\u6b63\u5e38"，\u7528\u65e0\u76d1\u7763\u6216\u534a\u76d1\u7763\u65b9\u6cd5\u5efa\u6a21"\u6b63\u5e38"\u6570\u636e\u7684\u5206\u5e03\u3002<\/p>\n

        2.7.2 \u5e38\u7528\u5f02\u5e38\u68c0\u6d4b\u65b9\u6cd5<\/h4>\n\n\u65b9\u6cd5\u539f\u7406\u9002\u7528\u573a\u666f\u4f18\u7f3a\u70b9<\/tr>\n\n\n\n\n\n\n\n
        Isolation Forest<\/td>\n\u968f\u673a\u5207\u5272\u7279\u5f81\u7a7a\u95f4，\u5f02\u5e38\u70b9\u66f4\u5bb9\u6613\u88ab\u9694\u79bb<\/td>\n\u9ad8\u7ef4\u8fde\u7eed\u7279\u5f81<\/td>\n\u5feb，\u4f46\u4e0d\u9002\u7528\u4e8e\u5c40\u90e8\u5f02\u5e38<\/td>\n<\/tr>\n
        LOF<\/td>\n\u57fa\u4e8e\u5c40\u90e8\u5bc6\u5ea6，\u5bc6\u5ea6\u4f4e\u4e8e\u90bb\u5c45\u5219\u4e3a\u5f02\u5e38<\/td>\n\u5bc6\u5ea6\u5dee\u5f02\u660e\u663e<\/td>\n\u8ba1\u7b97\u590d\u6742\u5ea6 O(n^2)<\/td>\n<\/tr>\n
        One-Class SVM<\/td>\n\u5728\u6838\u7a7a\u95f4\u4e2d\u7528\u4e00\u4e2a\u8d85\u7403\u9762\u5305\u88f9\u6b63\u5e38\u6570\u636e<\/td>\n\u4e2d\u7b49\u7ef4\u5ea6<\/td>\n\u5bf9\u6838\u53c2\u6570\u654f\u611f<\/td>\n<\/tr>\n
        Autoencoder<\/td>\n\u6b63\u5e38\u6837\u672c\u91cd\u5efa\u8bef\u5dee\u5c0f，\u5f02\u5e38\u6837\u672c\u91cd\u5efa\u8bef\u5dee\u5927<\/td>\n\u56fe\u50cf\u3001\u5e8f\u5217\u7b49\u975e\u7ed3\u6784\u5316\u6570\u636e<\/td>\n\u9700\u8981\u8db3\u591f\u591a\u7684\u6b63\u5e38\u6837\u672c<\/td>\n<\/tr>\n
        GANomaly<\/td>\n\u7528 GAN \u5b66\u4e60\u6b63\u5e38\u6837\u672c\u7684\u751f\u6210\u5206\u5e03<\/td>\n\u9ad8\u7ef4\u590d\u6742\u6570\u636e<\/td>\n\u8bad\u7ec3\u4e0d\u7a33\u5b9a<\/td>\n<\/tr>\n
        Deep SVDD<\/td>\n\u7528\u795e\u7ecf\u7f51\u7edc\u5c06\u6b63\u5e38\u6837\u672c\u6620\u5c04\u5230\u7d27\u51d1\u7684\u8d85\u7403\u9762\u5185<\/td>\n\u5927\u89c4\u6a21\u6570\u636e<\/td>\n\u9700\u8981\u7cbe\u7ec6\u8c03\u53c2<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

        from<\/span> sklearn.<\/span>ensemble import<\/span> IsolationForest<\/p>\n

        # Isolation Forest<\/span>
        \niso_forest =<\/span> IsolationForest(<\/span>
        \n contamination=<\/span>0.01<\/span>,<\/span>
        \n random_state=<\/span>42<\/span>,<\/span>
        \n n_estimators=<\/span>100<\/span>,<\/span>
        \n max_samples=<\/span>'auto'<\/span>
        \n)<\/span>
        \ny_pred =<\/span> iso_forest.<\/span>fit_predict(<\/span>X)<\/span> # \u6b63\u5e38=1, \u5f02\u5e38=-1<\/span><\/p>\n

        # One-Class SVM<\/span>
        \nfrom<\/span> sklearn.<\/span>svm import<\/span> OneClassSVM<\/p>\n

        oc_svm =<\/span> OneClassSVM(<\/span>
        \n nu=<\/span>0.01<\/span>,<\/span>
        \n kernel=<\/span>'rbf'<\/span>,<\/span>
        \n gamma=<\/span>'auto'<\/span>
        \n)<\/span>
        \ny_pred =<\/span> oc_svm.<\/span>fit_predict(<\/span>X)<\/span><\/p>\n

        # Autoencoder \u65b9\u6cd5<\/span>
        \nimport<\/span> torch.<\/span>nn as<\/span> nn<\/p>\n

        class<\/span> AnomalyAE<\/span>(<\/span>nn.<\/span>Module)<\/span>:<\/span>
        \n """\u7528\u91cd\u5efa\u8bef\u5dee\u68c0\u6d4b\u5f02\u5e38\u7684\u81ea\u7f16\u7801\u5668"""<\/span>
        \n def<\/span> __init__<\/span>(<\/span>self,<\/span> input_dim,<\/span> encoding_dim=<\/span>16<\/span>)<\/span>:<\/span>
        \n super<\/span>(<\/span>)<\/span>.<\/span>__init__(<\/span>)<\/span>
        \n self.<\/span>encoder =<\/span> nn.<\/span>Sequential(<\/span>
        \n nn.<\/span>Linear(<\/span>input_dim,<\/span> 64<\/span>)<\/span>,<\/span>
        \n nn.<\/span>ReLU(<\/span>)<\/span>,<\/span>
        \n nn.<\/span>Linear(<\/span>64<\/span>,<\/span> encoding_dim)<\/span>,<\/span>
        \n nn.<\/span>ReLU(<\/span>)<\/span>
        \n )<\/span>
        \n self.<\/span>decoder =<\/span> nn.<\/span>Sequential(<\/span>
        \n nn.<\/span>Linear(<\/span>encoding_dim,<\/span> 64<\/span>)<\/span>,<\/span>
        \n nn.<\/span>ReLU(<\/span>)<\/span>,<\/span>
        \n nn.<\/span>Linear(<\/span>64<\/span>,<\/span> input_dim)<\/span>,<\/span>
        \n nn.<\/span>Sigmoid(<\/span>)<\/span>
        \n )<\/span><\/p>\n

        def<\/span> forward<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
        \n return<\/span> self.<\/span>decoder(<\/span>self.<\/span>encoder(<\/span>x)<\/span>)<\/span><\/p>\n

        def<\/span> anomaly_score<\/span>(<\/span>self,<\/span> x)<\/span>:<\/span>
        \n reconstruction =<\/span> self.<\/span>forward(<\/span>x)<\/span>
        \n return<\/span> (<\/span>(<\/span>x –<\/span> reconstruction)<\/span> **<\/span> 2<\/span>)<\/span>.<\/span>mean(<\/span>dim=<\/span>1<\/span>)<\/span><\/p>\n

        # \u8bad\u7ec3：\u53ea\u7528"\u6b63\u5e38"\u6837\u672c<\/span>
        \nmodel =<\/span> AnomalyAE(<\/span>input_dim=<\/span>X_train.<\/span>shape[<\/span>1<\/span>]<\/span>)<\/span>
        \ncriterion =<\/span> nn.<\/span>MSELoss(<\/span>)<\/span>
        \noptimizer =<\/span> torch.<\/span>optim.<\/span>Adam(<\/span>model.<\/span>parameters(<\/span>)<\/span>)<\/span><\/p>\n

        for<\/span> epoch in<\/span> range<\/span>(<\/span>100<\/span>)<\/span>:<\/span>
        \n X_normal =<\/span> X_train[<\/span>y_train ==<\/span> 0<\/span>]<\/span>
        \n recon =<\/span> model(<\/span>X_normal)<\/span>
        \n loss =<\/span> criterion(<\/span>recon,<\/span> X_normal)<\/span>
        \n optimizer.<\/span>zero_grad(<\/span>)<\/span>
        \n loss.<\/span>backward(<\/span>)<\/span>
        \n optimizer.<\/span>step(<\/span>)<\/span><\/p>\n

        # \u63a8\u7406：\u91cd\u5efa\u8bef\u5dee\u5927\u7684\u5224\u4e3a\u5f02\u5e38<\/span>
        \nwith<\/span> torch.<\/span>no_grad(<\/span>)<\/span>:<\/span>
        \n scores =<\/span> model.<\/span>anomaly_score(<\/span>torch.<\/span>FloatTensor(<\/span>X_test)<\/span>)<\/span>
        \n threshold =<\/span> np.<\/span>percentile(<\/span>scores.<\/span>numpy(<\/span>)<\/span>,<\/span> 99<\/span>)<\/span>
        \n y_pred =<\/span> (<\/span>scores ><\/span> threshold)<\/span>.<\/span>int<\/span>(<\/span>)<\/span>.<\/span>numpy(<\/span>)<\/span><\/p>\n

        2.7.3 \u5f02\u5e38\u68c0\u6d4b\u4e0e\u5206\u7c7b\u65b9\u6cd5\u7684\u6839\u672c\u533a\u522b<\/h4>\n\n\u7ef4\u5ea6\u4f20\u7edf\u5206\u7c7b\u65b9\u6cd5\u5f02\u5e38\u68c0\u6d4b\u65b9\u6cd5<\/tr>\n\n\n\n\n\n\n\n
        \u8bad\u7ec3\u6570\u636e<\/td>\n\u9700\u8981\u6b63\u8d1f\u6837\u672c<\/td>\n\u53ea\u9700\u6b63\u5e38\u6837\u672c（\u65e0\u76d1\u7763）<\/td>\n<\/tr>\n
        \u95ee\u9898\u5b9a\u4e49<\/td>\n\u5b66\u4e60\u51b3\u7b56\u8fb9\u754c<\/td>\n\u5b66\u4e60\u6b63\u5e38\u6570\u636e\u7684\u5206\u5e03<\/td>\n<\/tr>\n
        \u5bf9\u4e0d\u5e73\u8861\u7684\u9002\u5e94<\/td>\n\u9700\u8981\u91c7\u6837\/\u52a0\u6743<\/td>\n\u5929\u7136\u9002\u5e94<\/td>\n<\/tr>\n
        \u8f93\u51fa<\/td>\n\u7c7b\u522b\u6807\u7b7e<\/td>\n\u5f02\u5e38\u5206\u6570 + \u9608\u503c<\/td>\n<\/tr>\n
        \u53ef\u89e3\u91ca\u6027<\/td>\n\u7279\u5f81\u91cd\u8981\u6027<\/td>\n\u91cd\u5efa\u8bef\u5dee \/ \u5bc6\u5ea6\u503c<\/td>\n<\/tr>\n
        \u9002\u7528\u4e0d\u5e73\u8861\u7387<\/td>\n\u4e00\u822c < 1:100<\/td>\n\u53ef > 1:10,000<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

        \u9762\u8bd5\u5b98\u8ffd\u95ee\u94fe<\/h4>\n

        Q：\u5f02\u5e38\u68c0\u6d4b\u65b9\u6cd5\u80fd\u66ff\u4ee3\u91c7\u6837\u65b9\u6cd5\u5417？<\/p>\n

        A： \u4e0d\u80fd\u5b8c\u5168\u66ff\u4ee3\u3002\u5f02\u5e38\u68c0\u6d4b\u9002\u7528\u4e8e\u6781\u4e0d\u5e73\u8861（1:1000+）\u4e14\u6ca1\u6709\u8db3\u591f\u7684\u6b63\u6837\u672c\u7684\u60c5\u51b5\u3002\u7ecf\u9a8c\u6cd5\u5219：\u6b63\u6837\u672c\u6570 < 100 \u65f6\u4f18\u5148\u8003\u8651\u5f02\u5e38\u68c0\u6d4b，\u6b63\u6837\u672c\u6570 > 1000 \u65f6\u4f18\u5148\u8003\u8651\u91c7\u6837 + \u5206\u7c7b\u3002<\/p>\n

        Q：\u4e3a\u4ec0\u4e48 Isolation Forest \u5728\u4e0d\u5e73\u8861\u6570\u636e\u4e0a\u8868\u73b0\u597d？<\/p>\n

        A： Isolation Forest \u7684\u6838\u5fc3\u5047\u8bbe\u662f"\u5f02\u5e38\u70b9\u66f4\u5bb9\u6613\u88ab\u9694\u79bb"\u2014\u2014\u5c11\u91cf\u5207\u5272\u5c31\u80fd\u5c06\u5f02\u5e38\u70b9\u5206\u79bb\u51fa\u6765\u3002\u591a\u6570\u7c7b\u6837\u672c（\u6b63\u5e38）\u5bc6\u96c6\u5206\u5e03\u5728\u7279\u5f81\u7a7a\u95f4\u4e2d，\u9700\u8981\u66f4\u591a\u5207\u5272\u624d\u80fd\u9694\u79bb；\u5c11\u6570\u7c7b\u6837\u672c（\u5f02\u5e38）\u7a00\u758f，\u66f4\u5bb9\u6613\u88ab\u9694\u79bb，\u56e0\u6b64\u5176\u8def\u5f84\u957f\u5ea6（\u4ece\u6839\u8282\u70b9\u5230\u53f6\u5b50\u8282\u70b9\u7684\u5207\u5272\u6b21\u6570）\u66f4\u77ed\u3002\u6240\u4ee5\u7528\u8def\u5f84\u957f\u5ea6\u4f5c\u4e3a\u5f02\u5e38\u5206\u6570\u5929\u7136\u9002\u7528\u4e8e\u4e0d\u5e73\u8861\u6570\u636e\u3002<\/p>\n

        Q：Autoencoder \u68c0\u6d4b\u5f02\u5e38\u7684\u6838\u5fc3\u5047\u8bbe\u662f\u4ec0\u4e48？\u5b83\u6709\u4ec0\u4e48\u9650\u5236？<\/p>\n

        A： \u6838\u5fc3\u5047\u8bbe\u662f：Autoencoder \u53ea\u5728\u6b63\u5e38\u6837\u672c\u4e0a\u8bad\u7ec3，\u5b66\u4f1a\u4e86"\u6b63\u5e38\u6a21\u5f0f"\u7684\u4f4e\u7ef4\u6d41\u5f62\u3002\u6b63\u5e38\u6837\u672c\u91cd\u5efa\u8bef\u5dee\u5c0f，\u800c\u5f02\u5e38\u6837\u672c\u4e0d\u5728\u8fd9\u4e2a\u6d41\u5f62\u4e0a，\u91cd\u5efa\u8bef\u5dee\u5927\u3002\u9650\u5236：（1）\u5982\u679c\u5f02\u5e38\u6837\u672c\u6070\u597d\u843d\u5728\u6b63\u5e38\u6d41\u5f62\u4e0a（\u5373\u5f02\u5e38\u770b\u8d77\u6765\u50cf\u6b63\u5e38），\u4e0d\u4f1a\u88ab\u68c0\u6d4b\u51fa\u6765；（2）Autoencoder \u53ef\u80fd\u5b66\u5230"\u6052\u7b49\u6620\u5c04"（\u4ec0\u4e48\u90fd\u4e0d\u538b\u7f29），\u5bfc\u81f4\u6240\u6709\u6837\u672c\u91cd\u5efa\u8bef\u5dee\u90fd\u5f88\u5c0f\u2014\u2014\u9700\u8981\u7528\u6b63\u5219\u5316\u6216\u964d\u566a\u81ea\u7f16\u7801\u5668\u6765\u907f\u514d\u3002<\/p>\n

        \n

        2.8 \u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406\u7684\u5b8c\u6574\u51b3\u7b56\u6d41\u7a0b<\/h3>\n

        \u9762\u5bf9\u4e00\u4e2a\u5b9e\u9645\u7684\u4e0d\u5e73\u8861\u95ee\u9898\u65f6，\u53ef\u4ee5\u6309\u4ee5\u4e0b\u51b3\u7b56\u6811\u8fdb\u884c\u9009\u62e9：<\/p>\n

        \u6570\u636e\u662f\u5426\u4e0d\u5e73\u8861？
        \n\u251c\u2500\u2500 \u4e0d\u5e73\u8861\u7387 < 1:10 \u2192 \u8f7b\u5ea6\u4e0d\u5e73\u8861
        \n\u2502 \u2514\u2500\u2500 \u8c03\u6574\u8bc4\u4f30\u6307\u6807 + \u8f7b\u5fae\u6743\u91cd\u8c03\u6574\u5373\u53ef
        \n\u2502
        \n\u251c\u2500\u2500 \u4e0d\u5e73\u8861\u7387 1:10 ~ 1:100 \u2192 \u4e2d\u5ea6\u4e0d\u5e73\u8861
        \n\u2502 \u251c\u2500\u2500 \u9996\u9009：SMOTE + \u51b3\u7b56\u6811 \/ XGBoost
        \n\u2502 \u251c\u2500\u2500 \u5907\u9009：Focal Loss（\u6df1\u5ea6\u6a21\u578b）\u6216\u4ee3\u4ef7\u654f\u611f\u5b66\u4e60
        \n\u2502 \u2514\u2500\u2500 \u6ce8\u610f：\u5c1d\u8bd5\u4e0d\u540c\u7684 SMOTE \u53d8\u4f53
        \n\u2502
        \n\u251c\u2500\u2500 \u4e0d\u5e73\u8861\u7387 1:100 ~ 1:10,000 \u2192 \u91cd\u5ea6\u4e0d\u5e73\u8861
        \n\u2502 \u251c\u2500\u2500 EasyEnsemble \/ Balanced Random Forest
        \n\u2502 \u251c\u2500\u2500 Focal Loss + \u6570\u636e\u589e\u5f3a
        \n\u2502 \u251c\u2500\u2500 \u4ee3\u4ef7\u654f\u611f + \u9608\u503c\u79fb\u52a8
        \n\u2502 \u2514\u2500\u2500 \u53ef\u4ee5\u5148\u5c1d\u8bd5\u964d\u7ef4 + SMOTE
        \n\u2502
        \n\u2514\u2500\u2500 \u4e0d\u5e73\u8861\u7387 > 1:10,000 \u2192 \u6781\u5ea6\u4e0d\u5e73\u8861
        \n \u251c\u2500\u2500 \u5f02\u5e38\u68c0\u6d4b\u65b9\u6cd5 (Isolation Forest \/ Autoencoder)
        \n \u251c\u2500\u2500 \u8f6c\u4e3a\u5f02\u5e38\u68c0\u6d4b\u95ee\u9898，\u653e\u5f03\u5206\u7c7b\u601d\u8def
        \n \u2514\u2500\u2500 \u5982\u679c\u4ecd\u6709\u6807\u7b7e：\u7528 EasyEnsemble \u4f46\u4e0d\u62b1\u592a\u5927\u671f\u671b<\/p>\n

        # \u4e00\u4e2a\u5b9e\u7528\u7684\u4e0d\u5e73\u8861\u6570\u636e\u5904\u7406\u6d41\u6c34\u7ebf\u793a\u4f8b<\/span>
        \nfrom<\/span> sklearn.<\/span>pipeline import<\/span> Pipeline
        \nfrom<\/span> sklearn.<\/span>preprocessing import<\/span> StandardScaler
        \nfrom<\/span> sklearn.<\/span>model_selection import<\/span> StratifiedKFold
        \nfrom<\/span> sklearn.<\/span>metrics import<\/span> classification_report<\/p>\n

        from<\/span> imblearn.<\/span>over_sampling import<\/span> SMOTE,<\/span> RandomOverSampler
        \nfrom<\/span> imblearn.<\/span>under_sampling import<\/span> RandomUnderSampler
        \nfrom<\/span> imblearn.<\/span>pipeline import<\/span> Pipeline as<\/span> ImbPipeline
        \nfrom<\/span> xgboost import<\/span> XGBClassifier<\/p>\n

        def<\/span> build_imbalanced_pipeline<\/span>(<\/span>ratio)<\/span>:<\/span>
        \n """\u6839\u636e\u4e0d\u5e73\u8861\u7387\u81ea\u52a8\u9009\u62e9\u5904\u7406\u7b56\u7565"""<\/span>
        \n if<\/span> ratio <<\/span> 10<\/span>:<\/span>
        \n sampler =<\/span> RandomOverSampler(<\/span>sampling_strategy=<\/span>'auto'<\/span>)<\/span>
        \n elif<\/span> ratio <<\/span> 100<\/span>:<\/span>
        \n sampler =<\/span> SMOTE(<\/span>sampling_strategy=<\/span>0.5<\/span>,<\/span> k_neighbors=<\/span>5<\/span>)<\/span>
        \n elif<\/span> ratio <<\/span> 10000<\/span>:<\/span>
        \n from<\/span> imblearn.<\/span>ensemble import<\/span> EasyEnsembleClassifier
        \n return<\/span> Pipeline(<\/span>[<\/span>
        \n (<\/span>'scaler'<\/span>,<\/span> StandardScaler(<\/span>)<\/span>)<\/span>,<\/span>
        \n (<\/span>'classifier'<\/span>,<\/span> EasyEnsembleClassifier(<\/span>
        \n n_estimators=<\/span>10<\/span>,<\/span>
        \n base_estimator=<\/span>XGBClassifier(<\/span>eval_metric=<\/span>'logloss'<\/span>)<\/span>,<\/span>
        \n random_state=<\/span>42<\/span>
        \n )<\/span>)<\/span>
        \n ]<\/span>)<\/span>
        \n else<\/span>:<\/span>
        \n from<\/span> sklearn.<\/span>ensemble import<\/span> IsolationForest
        \n contamination =<\/span> min<\/span>(<\/span>1.0<\/span> \/<\/span> (<\/span>1<\/span> +<\/span> ratio)<\/span>,<\/span> 0.5<\/span>)<\/span>
        \n return<\/span> Pipeline(<\/span>[<\/span>
        \n (<\/span>'scaler'<\/span>,<\/span> StandardScaler(<\/span>)<\/span>)<\/span>,<\/span>
        \n (<\/span>'classifier'<\/span>,<\/span> IsolationForest(<\/span>contamination=<\/span>contamination,<\/span> random_state=<\/span>42<\/span>)<\/span>)<\/span>
        \n ]<\/span>)<\/span><\/p>\n

        return<\/span> ImbPipeline(<\/span>[<\/span>
        \n (<\/span>'scaler'<\/span>,<\/span> StandardScaler(<\/span>)<\/span>)<\/span>,<\/span>
        \n (<\/span>'sampler'<\/span>,<\/span> sampler)<\/span>,<\/span>
        \n (<\/span>'classifier'<\/span>,<\/span> XGBClassifier(<\/span>
        \n scale_pos_weight=<\/span>ratio,<\/span>
        \n eval_metric=<\/span>'auc'<\/span>,<\/span>
        \n use_label_encoder=<\/span>False<\/span>
        \n )<\/span>)<\/span>
        \n ]<\/span>)<\/span><\/p>\n

        \n

        \u672c\u7ae0\u603b\u7ed3<\/h3>\n\n\u8003\u70b9\u6838\u5fc3\u8981\u70b9\u9762\u8bd5\u9891\u5ea6<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n
        \u51c6\u786e\u7387\u9677\u9631<\/td>\n\u4e0d\u5e73\u8861\u6570\u636e\u4e0b Accuracy \u5931\u6548；\u7528 Precision\/Recall\/F1\/AUC-PR \u66ff\u4ee3<\/td>\n\u2b50\u2b50\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        \u8fc7\u91c7\u6837 vs \u6b20\u91c7\u6837<\/td>\n\u8fc7\u91c7\u6837\u4fdd\u7559\u4fe1\u606f\u4f46\u6613\u8fc7\u62df\u5408；\u6b20\u91c7\u6837\u4fe1\u606f\u635f\u5931\u4f46\u8bad\u7ec3\u5feb<\/td>\n\u2b50\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        SMOTE \u539f\u7406<\/td>\n\u5728\u5c11\u6570\u7c7b\u8fd1\u90bb\u95f4\u63d2\u503c\u751f\u6210\u65b0\u6837\u672c；\u5173\u952e\u662f\u7406\u89e3\u63d2\u503c\u5047\u8bbe\u548c\u5c40\u9650<\/td>\n\u2b50\u2b50\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        SMOTE \u53d8\u4f53<\/td>\nBorderline\/ADASYN\/KMeansSMOTE\/SMOTE-ENN \u5404\u6709\u9002\u7528\u573a\u666f<\/td>\n\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        Focal Loss \u63a8\u5bfc<\/td>\n\u6807\u51c6 CE -> \u52a0\u5165 (1-p_t)^gamma \u8c03\u8282\u56e0\u5b50；gamma \u548c alpha \u7684\u5206\u5de5<\/td>\n\u2b50\u2b50\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        \u68af\u5ea6\u89e3\u91ca Focal Loss<\/td>\n\u6613\u5206\u6837\u672c\u68af\u5ea6\u88ab\u538b\u5236，\u96be\u5206\u6837\u672c\u68af\u5ea6\u4fdd\u6301\u2014\u2014\u52a8\u6001\u8c03\u8282<\/td>\n\u2b50\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        \u4ee3\u4ef7\u654f\u611f\u5b66\u4e60<\/td>\n\u6743\u91cd\u77e9\u9635\u3001\u9608\u503c\u79fb\u52a8\u3001scale_pos_weight \u4e09\u79cd\u5b9e\u73b0\u65b9\u5f0f<\/td>\n\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        EasyEnsemble<\/td>\n\u591a\u6b21\u6b20\u91c7\u6837 + \u57fa\u5206\u7c7b\u5668\u96c6\u6210，\u4fdd\u7559\u5168\u90e8\u591a\u6570\u7c7b\u4fe1\u606f<\/td>\n\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        BalanceCascade<\/td>\nEasyEnsemble + \u7ea7\u8054\u6dd8\u6c70\u5df2\u5206\u7c7b\u6b63\u786e\u7684\u591a\u6570\u7c7b<\/td>\n\u2b50\u2b50<\/td>\n<\/tr>\n
        \u5f02\u5e38\u68c0\u6d4b\u65b9\u6cd5<\/td>\nIsolation Forest \/ Autoencoder \u9002\u7528\u4e8e\u6781\u4e0d\u5e73\u8861\u573a\u666f<\/td>\n\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n
        \u51b3\u7b56\u6d41\u7a0b<\/td>\n\u6309\u4e0d\u5e73\u8861\u7387\u5206\u56db\u7ea7：\u8f7b\u5ea6->\u52a0\u6743\u3001\u4e2d\u5ea6->SMOTE\u3001\u91cd\u5ea6->\u96c6\u6210\u3001\u6781\u5ea6->\u5f02\u5e38\u68c0\u6d4b<\/td>\n\u2b50\u2b50\u2b50\u2b50<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
        \n

        \u5ef6\u4f38\u9605\u8bfb：<\/p>\n

          \n
        • He, H., & Garcia, E. A. (2009). \u201cLearning from Imbalanced Data.\u201d IEEE TKDE.<\/li>\n
        • Lin, T. Y., et al. (2017). \u201cFocal Loss for Dense Object Detection.\u201d ICCV.<\/li>\n
        • Liu, F. T., et al. (2008). \u201cIsolation Forest.\u201d ICDM.<\/li>\n
        • Chawla, N. V., et al. (2002). \u201cSMOTE: Synthetic Minority Over-sampling Technique.\u201d JAIR.<\/li>\n
        • \u9762\u8bd5\u52a0\u5206\u9879：\u4e86\u89e3\u5982\u4f55\u5904\u7406\u591a\u5206\u7c7b\u4e0d\u5e73\u8861（\u5982 Macro-F1）\u548c\u5728\u7ebf\u5b66\u4e60\u573a\u666f\u4e0b\u7684\u4e0d\u5e73\u8861\u95ee\u9898\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

          \u5f15\u5b50：\u4e09\u9053\u8ba9\u4f60\u6000\u7591\u4eba\u751f\u7684\u9762\u8bd5\u9898
          \n\u5728\u5f00\u59cb\u4e4b\u524d，\u8bf7\u5148\u601d\u8003\u4e0b\u9762\u51e0\u4e2a\u95ee\u9898\u3002\u5982\u679c\u4f60\u80fd\u4e00\u53e3\u6c14\u5168\u90e8\u7b54\u51fa\u6765，\u8bf4\u660e\u4f60\u5bf9\u6570\u636e\u4e0d\u5e73\u8861\u5df2\u7ecf\u6709\u4e86\u6bd4\u8f83\u900f\u5f7b\u7684\u7406\u89e3；\u5982\u679c\u7b54\u4e0d\u4e0a\u6765，\u90a3\u4e48\u8fd9\u4e00\u7ae0\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002\u95ee\u9898 1：\u4f60\u5728\u4e00\u4e2a\u4e8c\u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3\u6a21\u578b，\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387\u8fbe\u5230\u4e86 99.3%\u3002\u4f60\u975e\u5e38\u9ad8\u5174，\u4f46\u4f60\u7684\u4e0a\u7ea7\u770b\u4e86\u4e00\u773c\u6570\u636e\u5206\u5e03\u8bf4：\u201c\u6b63\u6837\u672c\u53ea\u5360 0.7%，\u4f60\u8fd9\u6a21\u578b\u4ec0\u4e48\u90fd\u6ca1\u5b66\u5230\u3002\u201d \u4e3a<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,207],"topic":[],"class_list":["post-89949","post","type-post","status-publish","format-standard","hentry","category-server","tag-50","tag-207"],"yoast_head":"\n\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/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:\/\/www.wsisp.com\/helps\/89949.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u5f15\u5b50：\u4e09\u9053\u8ba9\u4f60\u6000\u7591\u4eba\u751f\u7684\u9762\u8bd5\u9898 \u5728\u5f00\u59cb\u4e4b\u524d，\u8bf7\u5148\u601d\u8003\u4e0b\u9762\u51e0\u4e2a\u95ee\u9898\u3002\u5982\u679c\u4f60\u80fd\u4e00\u53e3\u6c14\u5168\u90e8\u7b54\u51fa\u6765，\u8bf4\u660e\u4f60\u5bf9\u6570\u636e\u4e0d\u5e73\u8861\u5df2\u7ecf\u6709\u4e86\u6bd4\u8f83\u900f\u5f7b\u7684\u7406\u89e3；\u5982\u679c\u7b54\u4e0d\u4e0a\u6765，\u90a3\u4e48\u8fd9\u4e00\u7ae0\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002\u95ee\u9898 1：\u4f60\u5728\u4e00\u4e2a\u4e8c\u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3\u6a21\u578b，\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387\u8fbe\u5230\u4e86 99.3%\u3002\u4f60\u975e\u5e38\u9ad8\u5174，\u4f46\u4f60\u7684\u4e0a\u7ea7\u770b\u4e86\u4e00\u773c\u6570\u636e\u5206\u5e03\u8bf4：\u201c\u6b63\u6837\u672c\u53ea\u5360 0.7%，\u4f60\u8fd9\u6a21\u578b\u4ec0\u4e48\u90fd\u6ca1\u5b66\u5230\u3002\u201d \u4e3a\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/89949.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-04T02:00:41+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"13 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/89949.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/89949.html\",\"name\":\"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-04T02:00:41+00:00\",\"dateModified\":\"2026-08-04T02:00:41+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/89949.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/89949.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/89949.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","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:\/\/www.wsisp.com\/helps\/89949.html","og_locale":"zh_CN","og_type":"article","og_title":"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u5f15\u5b50：\u4e09\u9053\u8ba9\u4f60\u6000\u7591\u4eba\u751f\u7684\u9762\u8bd5\u9898 \u5728\u5f00\u59cb\u4e4b\u524d，\u8bf7\u5148\u601d\u8003\u4e0b\u9762\u51e0\u4e2a\u95ee\u9898\u3002\u5982\u679c\u4f60\u80fd\u4e00\u53e3\u6c14\u5168\u90e8\u7b54\u51fa\u6765，\u8bf4\u660e\u4f60\u5bf9\u6570\u636e\u4e0d\u5e73\u8861\u5df2\u7ecf\u6709\u4e86\u6bd4\u8f83\u900f\u5f7b\u7684\u7406\u89e3；\u5982\u679c\u7b54\u4e0d\u4e0a\u6765，\u90a3\u4e48\u8fd9\u4e00\u7ae0\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002\u95ee\u9898 1：\u4f60\u5728\u4e00\u4e2a\u4e8c\u5206\u7c7b\u4efb\u52a1\u4e0a\u8bad\u7ec3\u6a21\u578b，\u6d4b\u8bd5\u96c6\u51c6\u786e\u7387\u8fbe\u5230\u4e86 99.3%\u3002\u4f60\u975e\u5e38\u9ad8\u5174，\u4f46\u4f60\u7684\u4e0a\u7ea7\u770b\u4e86\u4e00\u773c\u6570\u636e\u5206\u5e03\u8bf4：\u201c\u6b63\u6837\u672c\u53ea\u5360 0.7%，\u4f60\u8fd9\u6a21\u578b\u4ec0\u4e48\u90fd\u6ca1\u5b66\u5230\u3002\u201d \u4e3a","og_url":"https:\/\/www.wsisp.com\/helps\/89949.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-04T02:00:41+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"13 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/89949.html","url":"https:\/\/www.wsisp.com\/helps\/89949.html","name":"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-04T02:00:41+00:00","dateModified":"2026-08-04T02:00:41+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/89949.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/89949.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/89949.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u3010\u673a\u5668\u5b66\u4e60\u4e13\u680f\u30115.1 \u8bad\u7ec3\u5de5\u7a0b\uff1a\u6570\u636e\u4e0d\u5e73\u8861\u5904\u7406"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/89949","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=89949"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/89949\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=89949"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=89949"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=89949"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=89949"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}