{"id":943,"date":"2021-06-04T11:27:13","date_gmt":"2021-06-04T03:27:13","guid":{"rendered":"https:\/\/falbang.com\/?p=943"},"modified":"2021-06-04T11:27:14","modified_gmt":"2021-06-04T03:27:14","slug":"%e6%a8%a1%e5%9e%8b%e7%bb%93%e6%9e%9c%e5%8f%8d%e5%ae%9e%e9%99%85%ef%bc%8c%e5%8f%af%e8%83%bd%e5%bf%bd%e7%95%a5%e4%ba%86%e8%bf%99%e4%b8%80%e9%a1%b9%ef%bc%88%e9%99%84%e4%bb%a3%e7%a0%81%ef%bc%89","status":"publish","type":"post","link":"https:\/\/falbang.com\/?p=943","title":{"rendered":"\u6a21\u578b\u7ed3\u679c\u53cd\u5b9e\u9645\uff0c\u53ef\u80fd\u5ffd\u7565\u4e86\u8fd9\u4e00\u9879\uff08\u9644\u4ee3\u7801\uff09"},"content":{"rendered":"\n<p>\u5173\u6ce8\u201c\u91d1\u79d1\u5e94\u7528\u7814\u9662\u201d\uff0c\u56de\u590d\u201c\u793c\u5305\u201d\u9886\u53d6<strong>\u98ce\u63a7\u8d44\u6599\u5408\u96c6<\/strong><\/p>\n\n\n\n<p>\u6587\u672b\u6709\u60ca\u559c\u5c0f\u798f\u5229\uff0c\u8bb0\u5f97\u770b\u5230\u5e95\u5466<\/p>\n\n\n\n<ul><li>\u5728\u9884\u671f\u8303\u56f4\uff0c\u5076\u5c14\u51e0\u5468\u51fa\u73b0\u53d8\u91cf\u5206\u5e03\u4ea4\u53c9<\/li><li>\u4e0a\u7ebf\u6a21\u578b\u76d1\u63a7PSI\u67d0\u4e00\u5468\u7a81\u7136\u4e0a\u5347\uff0c\u4e0b\u4e00\u5468\u53c8\u5feb\u901f\u8dcc\u56de\u9884\u671f\u8303\u56f4<\/li><li>\u4e0a\u7ebf\u6a21\u578b\u6574\u4f53\u8868\u73b0\u9010\u6708\u7f13\u6162\u8870\u9000<\/li><\/ul>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"466\" src=\"https:\/\/falbang.com\/wp-content\/uploads\/2021\/06\/8-2-1024x466.jpg\" alt=\"\" class=\"wp-image-944\" srcset=\"https:\/\/falbang.com\/wp-content\/uploads\/2021\/06\/8-2-1024x466.jpg 1024w, https:\/\/falbang.com\/wp-content\/uploads\/2021\/06\/8-2-300x136.jpg 300w, https:\/\/falbang.com\/wp-content\/uploads\/2021\/06\/8-2-768x349.jpg 768w, https:\/\/falbang.com\/wp-content\/uploads\/2021\/06\/8-2-200x91.jpg 200w, https:\/\/falbang.com\/wp-content\/uploads\/2021\/06\/8-2.jpg 1080w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u5bfc\u81f4\u51fa\u73b0\u4e0a\u8ff0\u73b0\u8c61\u7684\u539f\u56e0\u6709\u5f88\u591a\uff0c\u6bd4\u5982\uff0c\u7b2c5\u70b9\u53ef\u80fd\u662f\u89c2\u5bdf\u5f02\u5e38\u7684\u51e0\u5468\u5185\u8fdb\u4ef6\u5ba2\u7fa4\u53d1\u751f\u4e86\u53d8\u5316\uff0c\u8fd9\u7c7b\u95ee\u9898\u53ef\u4ee5\u9759\u89c2\u5176\u53d8\u4e0d\u9700\u8981\u8c03\u6574\u6a21\u578b\u3002\u5f53\u7136\uff0c\u6709\u4e9b\u73b0\u8c61\u7684\u51fa\u73b0\u4f1a\u5bfc\u81f4\u6a21\u578b\u4e0a\u7ebf\u540e\u53d1\u751f\u6a21\u578b\u98ce\u9669\uff0c\u9700\u8981\u6a21\u578b\u5f00\u53d1\u4eba\u5458\u7acb\u523b\u89e3\u51b3\u3002<\/p>\n\n\n\n<p>\u5176\u4e2d\uff0c\u56de\u5f52\u7cfb\u6570\u7b26\u53f7\u4e0e\u5b9e\u9645\u60c5\u51b5\u76f8\u53cd\u8fd9\u79cd\u60c5\u51b5\uff0c\u5c31\u5c5e\u4e8e\u4e0a\u8ff0\u6a21\u578b\u98ce\u9669\u8303\u7574\u3002<\/p>\n\n\n\n<p><br>\u4e00\u822c\uff0c\u51fa\u73b0\u56de\u5f52\u7cfb\u6570\u7b26\u53f7\u4e0e\u5b9e\u9645\u60c5\u51b5\u76f8\u53cd\u662f\u56e0\u4e3a\u6a21\u578b\u53d8\u91cf\u4e2d\u5b58\u5728\u5171\u7ebf\u6027\u3002<\/p>\n\n\n\n<p>\u5171\u7ebf\u6027\u6307\u7684\u662f\u6a21\u578b\u7684\u53d8\u91cf\u4e4b\u95f4\u5b58\u5728\u8f83\u9ad8\u7684\u76f8\u5173\u6027\uff0c\u67d0\u4e00\u4e2a\u53d8\u91cf\u53ef\u4ee5\u88ab\u5176\u4ed6\u4e00\u90e8\u5206\u53d8\u91cf\u6240\u89e3\u91ca\u3002<\/p>\n\n\n\n<p><br>\u9053\u7406\u5f88\u7b80\u5355\uff1a\u4e00\u4e2a\u6307\u6807\u5149\u662f\u81ea\u5df1\u8868\u73b0\u7684\u8db3\u591f\u597d\u4e5f\u662f\u6ca1\u6709\u7528\u7684\uff0c\u8fd8\u8981\u8003\u8651\u5b83\u548c\u6240\u6709\u5176\u4ed6\u81ea\u53d8\u91cf\u4e4b\u95f4\u7684\u201c\u56e2\u961f\u534f\u4f5c\u80fd\u529b\u201d\u3002\u4e00\u4e2a\u4f18\u79c0\u7684\u56e2\u961f\uff0c\u5f80\u5f80\u9700\u8981\u961f\u5458\u4e4b\u95f4\u53d6\u957f\u8865\u77ed\uff0c\u5404\u81ea\u5206\u5de5\uff0c\u800c\u4e0d\u662f\u5927\u5bb6\u64c5\u957f\u7684\u4e8b\u7269\u90fd\u4e00\u6837\uff0c\u800c\u77ed\u677f\u5374\u65e0\u4eba\u5f25\u8865\u3002<\/p>\n\n\n\n<p><br>\u653e\u5728\u6a21\u578b\u4e0a\u4e5f\u662f\u4e00\u6837\u7684\u9053\u7406\u3002\u5982\u679c\u6a21\u578b\u4e2d\u7684\u81ea\u53d8\u91cf\u4e4b\u95f4\u5177\u6709\u5b8c\u5168\u591a\u91cd\u5171\u7ebf\u6027\uff0c\u90a3\u4e48\u8bad\u7ec3\u51fa\u6765\u7684\u7cfb\u6570\u4fbf\u4f1a\u5931\u53bb\u7edf\u8ba1\u5b66\u610f\u4e49\u3002\u5373\u4f7f\u662f\u4e0d\u5b8c\u5168\u5171\u7ebf\u6027\uff0c\u4e5f\u4f1a\u5bfc\u81f4\u7cfb\u6570\u5931\u771f\uff0c\u4ece\u800c\u5bfc\u81f4\u6a21\u578b\u7684\u6548\u679c\u65e0\u6cd5\u8fbe\u5230\u9884\u671f\u3002<\/p>\n\n\n\n<p><br>\u901a\u5e38\uff0c\u6211\u4eec\u53ef\u4ee5\u7528VIF\uff08\u65b9\u5dee\u81a8\u80c0\u7cfb\u6570\uff09\u6765\u68c0\u9a8c\u6a21\u578b\u5171\u7ebf\u6027\u95ee\u9898\uff0c\u5b8c\u6574Python\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import re\r\nimport os\r\nimport sys\r\nfrom pathlib import Path\r\nimport numpy as np\r\nimport pandas as pd\r\nimport matplotlib.pyplot as plt\r\nfrom termcolor import cprint\r\n\r\nimport lightgbm as lgb\r\nfrom joblib import Parallel, delayed\r\nimport time\r\n\r\nimport toad\r\nimport statsmodels.api as sm\r\nfrom statsmodels.stats.outliers_influence import variance_inflation_factor\r\nfrom sklearn.base import TransformerMixin, BaseEstimator, ClassifierMixin\r\nfrom sklearn.metrics import roc_auc_score, roc_curve\r\n\r\n\r\nclass Metrics:\r\n    @classmethod\r\n    def get_auc(cls, ytrue, yprob, **kwargs):\r\n        auc = roc_auc_score(ytrue, yprob)\r\n\r\n        if kwargs.get('symmetry', False) is True:\r\n            if auc &lt; 0.5:\r\n                auc = 1 - auc\r\n        return auc\r\n\r\n    @classmethod\r\n    def get_ks(cls, ytrue, yprob):\r\n        fpr, tpr, thr = roc_curve(ytrue, yprob)\r\n        ks = max(abs(tpr - fpr))\r\n        return ks\r\n\r\n    @classmethod\r\n    def get_gini(cls, ytrue, yprob, **kwargs):\r\n        auc = cls.get_auc(ytrue, yprob, **kwargs)\r\n        gini = 2 * auc - 1\r\n\r\n        return gini\r\n\r\n    @classmethod\r\n    def get_stat(cls, df_label, df_feature):\r\n        var = df_feature.name\r\n        df_data = pd.DataFrame({'val': df_feature, 'label': df_label})\r\n\r\n # statistics of total count, total ratio, bad count, bad rate\r\n        df_stat = df_data.groupby('val').agg(total=('label', 'count'),\r\n                                             bad=('label', 'sum'),\r\n                                             bad_rate=('label', 'mean'))\r\n        df_stat&#91;'var'] = var\r\n        df_stat&#91;'good'] = df_stat&#91;'total'] - df_stat&#91;'bad']\r\n        df_stat&#91;'total_ratio'] = df_stat&#91;'total'] \/ df_stat&#91;'total'].sum()\r\n        df_stat&#91;'good_density'] = df_stat&#91;'good'] \/ df_stat&#91;'good'].sum()\r\n        df_stat&#91;'bad_density'] = df_stat&#91;'bad'] \/ df_stat&#91;'bad'].sum()\r\n\r\n        eps = np.finfo(np.float32).eps\r\n        df_stat.loc&#91;:, 'iv'] = (df_stat&#91;'bad_density'] - df_stat&#91;'good_density']) * \\\r\n                               np.log((df_stat&#91;'bad_density'] + eps) \/ (df_stat&#91;'good_density'] + eps))\r\n\r\n        cols = &#91;'var', 'total', 'total_ratio', 'bad', 'bad_rate', 'iv', 'val']\r\n        df_stat = df_stat.reset_index()&#91;cols].set_index('var')\r\n        return df_stat\r\n\r\n    @classmethod\r\n    def get_iv(cls, df_label, df_feature):\r\n        df_stat = cls.get_stat(df_label, df_feature)\r\n        return df_stat&#91;'iv'].sum()\r\n\r\n\r\nclass Logit(BaseEstimator, ClassifierMixin):\r\n    def __init__(self):\r\n        self.model = None\r\n        self.detail = None\r\n\r\n    def fit(self, df_xtrain, df_ytrain, **kwargs):\r\n        df_xtrain_const = sm.add_constant(df_xtrain)\r\n\r\n# model training. default using newton method, if fail use bfgs method\r\n        try:\r\n            self.model = sm.Logit(df_ytrain, df_xtrain_const).fit(method='newton', maxiter=100)\r\n        except:\r\n            cprint(\"warning:  exist strong correlated features, \"\r\n                   \"got singular matrix in linear model, retry bfgs method instead.\",\r\n                   'red')\r\n            self.model = sm.Logit(df_ytrain, df_xtrain_const).fit(method='bfgs', maxiter=100)\r\n\r\n# prepare model result\r\n        self.detail = pd.DataFrame({'var': df_xtrain_const.columns.tolist(),\r\n                                    'coef': self.model.params,\r\n                                    'std_err': &#91;round(v, 3) for v in self.model.bse],\r\n                                    'z': &#91;round(v, 3) for v in self.model.tvalues],\r\n                                    'pvalue': &#91;round(v, 3) for v in self.model.pvalues]})\r\n        self.detail&#91;'std_data'] = df_xtrain.std()\r\n        self.detail&#91;'feature_importance'] = abs(self.detail&#91;'coef']) * self.detail&#91;'std_data']\r\n\r\n        return self\r\n\r\n    def predict(self, df_xtest, **kwargs):\r\n        return self.model.predict(sm.add_constant(df_xtest))\r\n\r\n    def predict_proba(self, df_xtest, **kwargs):\r\n        yprob = self.model.predict(sm.add_constant(df_xtest))\r\n        res = np.zeros((len(df_xtest), 2))\r\n        res&#91;:, 1] = yprob\r\n        res&#91;:, 0] = 1 - yprob\r\n        return res\r\n\r\n    def summary(self):\r\n        print(self.detail)\r\n\r\n    def get_importance(self):\r\n        return self.detail.drop('const', axis=0)\r\n\r\n\r\nclass VIFSelector(TransformerMixin):\r\n    def __init__(self):\r\n        self.detail = None\r\n        self.selected_features = list()\r\n\r\n    def fit(self, df_xtrain, df_ytrain, **kwargs):\r\n        vif_threshold = kwargs.get('vif_threshold', 10)\r\n        n_jobs = kwargs.get('n_jobs', -1)\r\n        feature_list = sorted(kwargs.get('feature_list', df_xtrain.columns.tolist()))\r\n\r\n # compute VIF\r\n        vif = Parallel(n_jobs=n_jobs)(\r\n            delayed(variance_inflation_factor)(df_xtrain.values, i) for i in range(df_xtrain&#91;feature_list].shape&#91;1]))\r\n\r\n # select features with VIF &lt; vif_threshold\r\n        df_res = pd.DataFrame({'var': feature_list,\r\n                               'vif': &#91;round(v, 3) for v in vif]})\r\n        df_res.loc&#91;:, 'selected'] = df_res&#91;'vif'] &lt; vif_threshold\r\n\r\n        self.detail = df_res\r\n        self.selected_features = df_res.loc&#91;(df_res&#91;'selected'] == True), 'var'].tolist()\r\n\r\n        return self\r\n\r\n    def transform(self, df_xtest, **kwargs):\r\n        feature_list = kwargs.get('feature_list', df_xtest.columns.tolist())\r\n        feature_list = sorted(set(feature_list) &amp; set(self.selected_features))\r\n        return df_xtest&#91;feature_list]\r\n\r\n    def summary(self):\r\n        print('selected features:')\r\n        print(self.selected_features)\r\n        print('summary')\r\n        print(self.detail)\r\n\r\n\r\nmodel_vif = VIFSelector()\r\nmodel_vif.fit(df_xtrain, df_ytrain)\r\n\r\nprint\uff08model_vif.detail\uff09<\/code><\/pre>\n\n\n\n<p>VIF\u9608\u503c\u9009\u53d6\u6807\u51c6\u76f8\u5bf9\u4e3b\u89c2\u6027\u3002\u901a\u5e38\uff0c\u5982\u679cVIF\u7684\u9608\u503c\u5c0f\u4e8e10\uff08\u4e25\u683c\u662f3\uff09\uff0c\u5219\u8bf4\u660e\u6a21\u578b\u6ca1\u6709\u591a\u91cd\u5171\u7ebf\u6027\u95ee\u9898\uff0c\u6a21\u578b\u6784\u5efa\u826f\u597d\uff1b\u53cd\u4e4b\u82e5VIF\u5927\u4e8e10\u8bf4\u660e\u6a21\u578b\u6784\u5efa\u8f83\u5dee\u3002<br>\u5f53\u9009\u51faVIF>3\u7684\u53d8\u91cf\u7279\u5f81\u540e\uff0c\u6709\u4ec0\u4e48\u65b9\u6cd5\u53ef\u4ee5\u79fb\u9664\u5171\u7ebf\u6027\u53d8\u91cf\u5462\uff1f<\/p>\n\n\n\n<p><br>\u8fd9\u91cc\u63d0\u4f9b\u4e00\u4e2a\u65b9\u6cd5\uff1a<strong>\u7efc\u5408\u8003\u8651\u5355\u53d8\u91cf\u6216\u591a\u53d8\u91cf\u7684AR\u503c\u3002<\/strong><\/p>\n\n\n\n<p><br>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c\u53d8\u91cfA\u540c\u65f6\u548c\u53d8\u91cfB\u3001\u53d8\u91cfC\u5b58\u5728\u8f83\u5f3a\u76f8\u5173\u6027\uff0c\u800c\u53d8\u91cfB\u548c\u53d8\u91cfC\u76f8\u5173\u6027\u8f83\u4f4e\uff0c\u90a3\u4e48\u7a76\u7adf\u5e94\u8be5\u4fdd\u7559A\u8fd8\u662f\u540c\u65f6\u4fdd\u7559B\u548cC\u5462\uff1f<\/p>\n\n\n\n<p><br>\u6709\u4e00\u4e2a\u6bd4\u8f83\u7b80\u5355\u7684\u601d\u8def\uff0c\u53ef\u4ee5\u8ba1\u7b97\u53d8\u91cfA\u7684\u5355\u53d8\u91cfAR\u503c\uff0c\u540c\u65f6\u7528\u53d8\u91cfB\u548c\u53d8\u91cfC\u8bad\u7ec3\u4e00\u4e2a\u7b80\u6613\u7684\u903b\u8f91\u56de\u5f52\u6a21\u578b\uff0c\u5e76\u8ba1\u7b97\u8fd9\u4e2a\u6a21\u578b\u7684AR\u503c\uff0c\u901a\u8fc7\u6bd4\u8f83\u53d8\u91cfA\u548c\u6a21\u578b\u7684AR\u503c\u6765\u5224\u65ad\u5230\u5e95\u5e94\u8be5\u4fdd\u7559A\u8fd8\u662fB+C\u3002<\/p>\n\n\n\n<p><br>\u5982\u679c\u53d8\u91cfA\u7684AR\u503c\u6bd4\u6a21\u578b\u7684AR\u503c\u8fd8\u9ad8\uff0c\u5c31\u8bf4\u660e\u53d8\u91cfA\u7684\u8868\u73b0\u4f1a\u4f18\u4e8eB\u548cC\u7684\u7efc\u5408\u8868\u73b0\uff0c\u8fd9\u65f6\uff0c\u5c31\u53ef\u4ee5\u5254\u9664\u6389B\u548cC\u4e86\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4f60\u559c\u6b22\u3001\u60f3\u8981\u770b\u66f4\u591a\u7684\u5e72\u8d27\u7c7b\u578b\u7684\u6587\u7ae0\uff0c\u53ef\u4ee5\u5173\u6ce8\u516c\u4f17\u53f7\u3010\u91d1\u79d1\u5e94\u7528\u7814\u9662\u3011\u5e76\u8bbe\u4e3a\u661f\u6807\uff0c\u987a\u4fbf\u8f6c\u53d1\u5206\u4eab\uff5e<\/p>\n\n\n\n<p><br>\u611f\u8c22\u60a8\u770b\u5230\u8fd9\u91cc\u5fae\u4fe1\u516c\u4f17\u53f7\u3010\u91d1\u79d1\u5e94\u7528\u7814\u9662\u3011\u5bf9\u8bdd\u6846\u56de\u590d<strong>\u201c\u5c0f\u798f\u5229\u201d<\/strong>\uff0c\u9886\u53d6\u7c89\u4e1d\u4e13\u5c5e\u4f18\u60e0\u5238\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img src=\"https:\/\/falbang.com\/wp-content\/uploads\/2021\/04\/%E6%B5%8B%E8%AF%951.png\" alt=\"\" class=\"wp-image-725\"\/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u5173\u6ce8\u201c\u91d1\u79d1\u5e94\u7528\u7814\u9662\u201d\uff0c\u56de\u590d\u201c\u793c\u5305\u201d\u9886\u53d6\u98ce\u63a7\u8d44\u6599\u5408\u96c6 \u6587\u672b\u6709\u60ca\u559c\u5c0f\u798f\u5229\uff0c\u8bb0\u5f97\u770b\u5230\u5e95\u5466 \u5728\u9884\u671f\u8303\u56f4\uff0c\u5076\u5c14\u51e0\u5468\u51fa\u73b0\u53d8 &hellip; <\/p>\n<p class=\"link-more\"><a href=\"https:\/\/falbang.com\/?p=943\" class=\"more-link\">Continue reading<span class=\"screen-reader-text\"> &#8220;\u6a21\u578b\u7ed3\u679c\u53cd\u5b9e\u9645\uff0c\u53ef\u80fd\u5ffd\u7565\u4e86\u8fd9\u4e00\u9879\uff08\u9644\u4ee3\u7801\uff09&#8221;<\/span><\/a><\/p>\n","protected":false},"author":182,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[32],"tags":[50,51],"_links":{"self":[{"href":"https:\/\/falbang.com\/index.php?rest_route=\/wp\/v2\/posts\/943"}],"collection":[{"href":"https:\/\/falbang.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/falbang.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/falbang.com\/index.php?rest_route=\/wp\/v2\/users\/182"}],"replies":[{"embeddable":true,"href":"https:\/\/falbang.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=943"}],"version-history":[{"count":1,"href":"https:\/\/falbang.com\/index.php?rest_route=\/wp\/v2\/posts\/943\/revisions"}],"predecessor-version":[{"id":945,"href":"https:\/\/falbang.com\/index.php?rest_route=\/wp\/v2\/posts\/943\/revisions\/945"}],"wp:attachment":[{"href":"https:\/\/falbang.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=943"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/falbang.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=943"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/falbang.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=943"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}