作者：杰少

Null Importance特征筛选

简介

目前数据量越来越大，数据特征维度也越来越高，这不仅对我们的计算存储带来了较大的挑战，与此同时，还会对模型的效果带来较大的损益。

• 如何既能节省内存计算资源，同时能拿到模型的提效是我们非常关心的一个问题。

本文我们介绍一种特征筛选策略 -- Null Importance 特征筛选策略，该策略在95%的数据竞赛中基本都可以拿到效果，带来不错的提升。

Null Importance

1. 核心思想

Null Importance的核心思想在于：

1. 计算不靠谱的特征重要性；

• 我们对标签进行随机shuffle，并计算特征重要性，这些特征重要性是“错误的”；

• 计算靠谱的特征重要性；

• • 对原始数据进行训练并得到特征重要性，这些特征重要性是“正确的”；

  1. 计算靠谱的特征重要性和不靠谱的特征重要性的差距/偏离度（自行设计Score函数）；

  2. 按照计算得到的Score进行批量的特征筛选，并计算线下验证分数；

  3. 选用线下分数最好的一些特征作为最终的特征；

  2. 实现步骤

  Null Importance算法的实现步骤为：

  1. 在原始数据集上运行模型并且记录每个特征重要性。以此作为基准；

  2. 构建Null importances分布：对我们的标签进行随机Shuffle，并且计算shuffle之后的特征的重要性；

  3. 对2进行多循环操作，得到多个不同shuffle之后的特征重要性；

  4. 设计score函数，得到未shuffle的特征重要性与shuffle之后特征重要性的偏离度，并以此设计特征筛选策略；

  5. 计算不同筛选情况下的模型的分数，并进行记录；

  6. 将分数最好的几个分数对应的特征进行返回。

  代码

  

  代码摘自：https://www.kaggle.com/ogrellier/feature-selection-with-null-importances。

  1. 特征重要性获取函数

  def get_feature_importances(data, shuffle, seed=None):# Gather real featurestrain_features = [f for f in data if f not in ['TARGET', 'SK_ID_CURR']]# Go over fold and keep track of CV score (train and valid) and feature importances# Shuffle target if requiredy = data['TARGET'].copy()if shuffle:# Here you could as well use a binomial distributiony = data['TARGET'].copy().sample(frac=1.0)# Fit LightGBM in RF mode, yes it's quicker than sklearn RandomForestdtrain = lgb.Dataset(data[train_features], y, free_raw_data=False, silent=True)lgb_params = {'objective': 'binary','boosting_type': 'rf','subsample': 0.623,'colsample_bytree': 0.7,'num_leaves': 127,'max_depth': 8,'seed': seed,'bagging_freq': 1,'n_jobs': 4}# Fit the modelclf = lgb.train(params=lgb_params, train_set=dtrain, num_boost_round=200, categorical_feature=categorical_feats)# Get feature importancesimp_df = pd.DataFrame()imp_df["feature"] = list(train_features)imp_df["importance_gain"] = clf.feature_importance(importance_type='gain')imp_df["importance_split"] = clf.feature_importance(importance_type='split')imp_df['trn_score'] = roc_auc_score(y, clf.predict(data[train_features]))return imp_df
  

  2.获取原始版本的特征重要性

  # Seed the unexpected randomness of this world
  np.random.seed(123)
  # Get the actual importance, i.e. without shuffling
  actual_imp_df = get_feature_importances(data=data, shuffle=False)
  

  3.获取多个target shuffle版本的特征重要性

  null_imp_df = pd.DataFrame()
  nb_runs = 80
  import time
  start = time.time()
  dsp = ''
  for i in range(nb_runs):# Get current run importancesimp_df = get_feature_importances(data=data, shuffle=True)imp_df['run'] = i + 1 # Concat the latest importances with the old onesnull_imp_df = pd.concat([null_imp_df, imp_df], axis=0)# Erase previous messagefor l in range(len(dsp)):print('\b', end='', flush=True)# Display current run and time usedspent = (time.time() - start) / 60dsp = 'Done with %4d of %4d (Spent %5.1f min)' % (i + 1, nb_runs, spent)print(dsp, end='', flush=True)
  

  4.计算Score

  4.1 Score计算方式1

  • 以未进行特征shuffle的特征重要性除以shuffle之后的0.75分位数作为我们的score；

  feature_scores = []
  for _f in actual_imp_df['feature'].unique():f_null_imps_gain = null_imp_df.loc[null_imp_df['feature'] == _f, 'importance_gain'].valuesf_act_imps_gain = actual_imp_df.loc[actual_imp_df['feature'] == _f, 'importance_gain'].mean()gain_score = np.log(1e-10 + f_act_imps_gain / (1 + np.percentile(f_null_imps_gain, 75)))  # Avoid didvide by zerof_null_imps_split = null_imp_df.loc[null_imp_df['feature'] == _f, 'importance_split'].valuesf_act_imps_split = actual_imp_df.loc[actual_imp_df['feature'] == _f, 'importance_split'].mean()split_score = np.log(1e-10 + f_act_imps_split / (1 + np.percentile(f_null_imps_split, 75)))  # Avoid didvide by zerofeature_scores.append((_f, split_score, gain_score))scores_df = pd.DataFrame(feature_scores, columns=['feature', 'split_score', 'gain_score'])

  4.2 计算方式2

  • shuffle target之后特征重要性低于实际target对应特征的重要性0.25分位数的次数百分比。

  correlation_scores = []
  for _f in actual_imp_df['feature'].unique():f_null_imps = null_imp_df.loc[null_imp_df['feature'] == _f, 'importance_gain'].valuesf_act_imps = actual_imp_df.loc[actual_imp_df['feature'] == _f, 'importance_gain'].valuesgain_score = 100 * (f_null_imps < np.percentile(f_act_imps, 25)).sum() / f_null_imps.sizef_null_imps = null_imp_df.loc[null_imp_df['feature'] == _f, 'importance_split'].valuesf_act_imps = actual_imp_df.loc[actual_imp_df['feature'] == _f, 'importance_split'].valuessplit_score = 100 * (f_null_imps < np.percentile(f_act_imps, 25)).sum() / f_null_imps.sizecorrelation_scores.append((_f, split_score, gain_score))corr_scores_df = pd.DataFrame(correlation_scores, columns=['feature', 'split_score', 'gain_score'])

  5. 计算特征筛选之后的最佳分数并记录相应特征

  • 选用筛选之后分数最好的特征作为最终特征即可。

  def score_feature_selection(df=None, train_features=None, cat_feats=None, target=None):# Fit LightGBM dtrain = lgb.Dataset(df[train_features], target, free_raw_data=False, silent=True)lgb_params = {'objective': 'binary','boosting_type': 'gbdt','learning_rate': .1,'subsample': 0.8,'colsample_bytree': 0.8,'num_leaves': 31,'max_depth': -1,'seed': 13,'n_jobs': 4,'min_split_gain': .00001,'reg_alpha': .00001,'reg_lambda': .00001,'metric': 'auc'}# Fit the modelhist = lgb.cv(params=lgb_params, train_set=dtrain, num_boost_round=2000,categorical_feature=cat_feats,nfold=5,stratified=True,shuffle=True,early_stopping_rounds=50,verbose_eval=0,seed=17)# Return the last mean / std values return hist['auc-mean'][-1], hist['auc-stdv'][-1]# features = [f for f in data.columns if f not in ['SK_ID_CURR', 'TARGET']]
  # score_feature_selection(df=data[features], train_features=features, target=data['TARGET'])for threshold in [0, 10, 20, 30 , 40, 50 ,60 , 70, 80 , 90, 95, 99]:split_feats     = [_f for _f, _score, _ in correlation_scores if _score >= threshold]split_cat_feats = [_f for _f, _score, _ in correlation_scores if (_score >= threshold) & (_f in categorical_feats)]gain_feats     = [_f for _f, _, _score in correlation_scores if _score >= threshold]gain_cat_feats = [_f for _f, _, _score in correlation_scores if (_score >= threshold) & (_f in categorical_feats)]print('Results for threshold %3d' % threshold)split_results = score_feature_selection(df=data, train_features=split_feats, cat_feats=split_cat_feats, target=data['TARGET'])print('\t SPLIT : %.6f +/- %.6f' % (split_results[0], split_results[1]))gain_results = score_feature_selection(df=data, train_features=gain_feats, cat_feats=gain_cat_feats, target=data['TARGET'])print('\t GAIN  : %.6f +/- %.6f' % (gain_results[0], gain_results[1]))

  适用问题

  

  Null Importance特征筛选策略是数据实践中必备的技能之一，该方法基本可以在95%以上的数据竞赛或者实践项目中都能取得一定的收益。

  参考文献

  

  1. https://www.kaggle.com/ogrellier/feature-selection-with-null-importances

  2. https://academic.oup.com/bioinformatics/article/26/10/1340/193348

  
  往期精彩回顾适合初学者入门人工智能的路线及资料下载机器学习及深度学习笔记等资料打印机器学习在线手册深度学习笔记专辑《统计学习方法》的代码复现专辑
  AI基础下载机器学习的数学基础专辑黄海广老师《机器学习课程》课件合集
  本站qq群851320808，加入微信群请扫码：