https://tianchi.aliyun.com/forum/postDetail?spm=5176.12586969.1002.6.6b17329eVBTIwN&postId=333161

    时序问题 - 图1

    • 时间特征
      年、月、日、小时、星期、季度、是否工作日等。

      1. def get_time_feature(df, col, keep=False):
      2.   """
      3.   为df增加时间特征列,包括:年,月,日,小时,dayofweek,weekofyear
      4.   :param df:
      5.   :param col: 时间列的列名
      6.   :param keep: 是否保留原始时间列
      7.   :return:
      8.   """
      9.   df_copy = df.copy()
      10.   prefix = col + "_"
      11.   df_copy[col] = pd.to_datetime(df_copy[col])
      12.   df_copy[prefix + 'year'] = df_copy[col].dt.year
      13.   df_copy[prefix + 'month'] = df_copy[col].dt.month
      14.   df_copy[prefix + 'day'] = df_copy[col].dt.day
      15.   df_copy[prefix + 'hour'] = df_copy[col].dt.hour
      16.   df_copy[prefix + 'weekofyear'] = df_copy[col].dt.weekofyear
      17.   df_copy[prefix + 'dayofweek'] = df_copy[col].dt.dayofweek
      18.   df_copy[prefix + 'is_wknd'] = df_copy[col].dt.dayofweek // 4
      19.   df_copy[prefix + 'quarter'] = df_copy[col].dt.quarter
      20.   df_copy[prefix + 'is_month_start'] = df_copy[col].dt.is_month_start.astype(int)
      21.   df_copy[prefix + 'is_month_end'] = df_copy[col].dt.is_month_end.astype(int)
      22.   if keep:
      23.       return df_copy
      24.   else:
      25.       return df_copy.drop([col], axis=1)
      26. df = get_time_feature(df, "time_col")

    • lag特征
      表示同一元素历史时间点的值(例如在这个赛题中，同一个unit昨天、前天、上周对应的使用量)

      1. keys = ['unit']
      2. val = 'qty'
      3. lag = 1
      4. df.groupby(keys)[val].transform(lambda x: x.shift(lag))

    • 滑动窗口统计特征：历史时间窗口内的统计值

      1. keys = ['unit']
      2. val = 'qty'
      3. window = 7
      4. df.groupby(keys)[val].transform(  
      5. lambda x: x.rolling(window=window, min_periods=3, win_type="triang").mean())
      6. df.groupby(keys)[val].transform(  
      7. lambda x: x.rolling(window=window, min_periods=3).std())

    • 指数加权移动平均

      1. keys = ['unit']
      2. val = 'qty'
      3. lag = 1
      4. alpha=0.95
      5. df_temp.groupby(keys)[val].transform(lambda x: x.shift(lag).ewm(alpha=alpha).mean())

    • 标签的转化
      当label的分布范围差异较大时，可以尝试做变换后再进行训练，训练完模型后，对测试集预测后再逆操作回来。
      常用的方法有log1p，适用于label值均为正值的场景。转化后的训练结果会更加稳定。

      1. # 训练时转化
      2. assert(train[target].min() >= 0)
      3. train[target] = np.log1p(train[target])
      4. # 预测结果需要逆操作
      5. pred = np.expm1(pred)

    • 验证集的切分
      时序预测问题验证集的切分常常采用按时间切分的方式，可以用sklearn中的TimeSeriesSplit。在验证集上获得模型的最佳迭代轮数之后，再用全量数据重新训练。重新训练的时候，迭代轮次可以是之前迭代轮次的k倍。k的参考值=全量数据样本量/除去验证集的数据样本量。
      参考如下代码：

      1. from sklearn.model_selection import TimeSeriesSplit
      2. import gc
      3. ts_folds = TimeSeriesSplit(n_splits = 5)
      4. for fold_n, (train_index, valid_index) in enumerate(ts_folds.split(train[used_features])):
      5. if fold_n in [0, 1, 2, 3]:  
      6.       continue  
      8. print('Training with validation') 
      9. trn_data = lgb.Dataset(train[used_features].iloc[train_index], label=train[target].iloc[train_index],  
      10. categorical_feature="")  
      11. val_data = lgb.Dataset(train[used_features].iloc[valid_index], label=train[target].iloc[valid_index],  
      12. categorical_feature="")  
      13. clf = lgb.train(params, trn_data, num_boost_round=N_round, valid_sets=[trn_data, val_data], verbose_eval=Verbose,  
      14. early_stopping_rounds=Early_Stopping_Rounds)
      15. val = clf.predict(train[used_features].iloc[valid_index])  
      16. if target_log:  
      17.       mae_ = mean_absolute_error(np.expm1(train.iloc[valid_index][target]), np.expm1(val))  
      18. else:  
      19.       mae_ = mean_absolute_error(train.iloc[valid_index][target], val)  
      21. print('MAE: {}'.format(mae_))  
      22. MAEs.append(mae_)  
      24. print("ReTraining on all data")  
      25. gc.enable()  
      26. del trn_data, val_data  
      27.   gc.collect()  
      28. Best_iteration = clf.best_iteration  
      29.   print("Best_iteration: ", Best_iteration)  
      30. trn_data = lgb.Dataset(train[used_features], label=train[target], categorical_feature="")  
      31. clf = lgb.train(params, trn_data, num_boost_round=int(Best_iteration * 1.2),  
      32. valid_sets=[trn_data], verbose_eval=Verbose)  
      33. pred = clf.predict(test[used_features])