半自動構造新特徵

半自動構造新特徵

主要原理是通過groupby(C1).agg(func)[N1]的方式來創造新特徵，實現特徵交叉
還需配合特徵篩選使用有更好效果
需要傳入gby_cols 分組列，stati_cols 統計列，func_list 統計函數
prefix_list的特徵前綴可自行修改
簡單易懂，無多餘功能，網上沒看到類似函數，自行寫一個

# 半自動構造新特徵
class FeatureCombination(object):
    def __init__(self):
        # init
        self.prefix_list = []

    def transform(self, X):
    	# 下面這三行爲可更改
        gby_cols = ['channel_1_name', 'customer_province', 'city', 'online_work_status', 'if_introduced', 'prepare_lessons']
        stati_cols = [['if_pay'], ['teacher_join_days'], ['pay_student_nums'], ['age'], ['course_nums'], ['mean_score'], ['mean_price']]
        func_list = [['mean']]
        combination_list = []

        for i in gby_cols:
            combination_list.append([i])
            item_list = [i]
            for j in gby_cols:
                if i !=j:
                    item_list.append(j)
                    combination_list.append(item_list)
                    combination_list.append([i, j])

        # 完成去重操作
        combination_list = np.unique([sorted(x) for x in combination_list]).tolist()
        # 完成特徵衍生
        for gby_col in combination_list:
            for stati_col in stati_cols:
                for func in func_list:
                    X = self.feature_combination(X, gby_col, stati_col, func)
        return  X
    
    # X 輸入dataframe, gby_cols 分組列, stati_cols 統計列, func_list 處理函數list
    def feature_combination(self, X, gby_cols, stati_col, func_list):
        gp = X[gby_cols+stati_col].groupby(gby_cols)[stati_col].agg(func_list).reset_index()

        gp_values = gp[','.join(stati_col)]
        prefix = "-".join(gby_cols+stati_col)
        prefix_list = [x +'_by_'+ prefix for x in func_list]

        # 將組合衍生後的特徵名稱保存起來
        self.prefix_list.append(prefix_list[0])

        res = pd.concat([gp[gby_cols], gp_values], axis=1, ignore_index=True)
        res.columns = gby_cols+prefix_list
        return pd.merge(X, res, how='left', on=gby_cols)