在快速训练时增加余票变化特征

data_preprocess.py

@@ -847,9 +847,12 @@ def preprocess_data_simple(df_input, is_train=False):
     df_input = df_input[df_input['hours_until_departure'] <= 480]
     df_input = df_input[df_input['baggage'] == 30]
 
-    # 保留真实的而不是补齐的数据
+    # 在hours_until_departure 的末尾 保留真实的而不是补齐的数据
     if not is_train:
-        df_input = df_input[df_input['is_filled'] == 0]
+        _tail_filled = df_input.groupby(['gid', 'baggage'])['is_filled'].transform(
+            lambda s: s.iloc[::-1].cummin().iloc[::-1]
+        )
+        df_input = df_input[~((df_input['is_filled'] == 1) & (_tail_filled == 1))]
     
     # 计算价格变化量
     df_input['price_change_amount'] = (
@@ -878,6 +881,18 @@ def preprocess_data_simple(df_input, is_train=False):
     
     # 可选：删除临时列
     df_input = df_input.drop(columns=['price_change_segment'])
+
+    # 仅在价格变化点记录余票变化量；其它非价格变化点置空（NaN） 
+    _price_diff = df_input.groupby(['gid', 'baggage'], group_keys=False)['adult_total_price'].diff()
+    _price_changed = _price_diff.notna() & _price_diff.ne(0)
+    _seats_diff = df_input.groupby(['gid', 'baggage'], group_keys=False)['seats_remaining'].diff()
+    df_input['seats_remaining_change_amount'] = _seats_diff.where(_price_changed).round(0)
+    # 前向填充 并 填充缺失值为0
+    df_input['seats_remaining_change_amount'] = (
+        df_input.groupby(['gid', 'baggage'], group_keys=False)['seats_remaining_change_amount']
+        .ffill()
+        .fillna(0)
+    )
     
     adult_price = df_input.pop('Adult_Total_Price')
     hours_until = df_input.pop('Hours_Until_Departure')
@@ -887,7 +902,7 @@ def preprocess_data_simple(df_input, is_train=False):
 
     # 训练过程
     if is_train:
-        df_target = df_input[(df_input['hours_until_departure'] >= 18) & (df_input['hours_until_departure'] <= 54)].copy()
+        df_target = df_input[(df_input['hours_until_departure'] >= 12) & (df_input['hours_until_departure'] <= 60)].copy()
         df_target = df_target.sort_values(
             by=['gid', 'hours_until_departure'],
             ascending=[True, False]
@@ -897,6 +912,7 @@ def preprocess_data_simple(df_input, is_train=False):
         prev_pct = df_target.groupby('gid', group_keys=False)['price_change_percent'].shift(1)
         prev_amo = df_target.groupby('gid', group_keys=False)['price_change_amount'].shift(1)
         prev_dur = df_target.groupby('gid', group_keys=False)['price_duration_hours'].shift(1)
+        prev_seats_amo = df_target.groupby('gid', group_keys=False)['seats_remaining_change_amount'].shift(1)
         drop_mask = (prev_pct > 0) & (df_target['price_change_percent'] < 0)
         
         df_drop_nodes = df_target.loc[drop_mask, ['gid', 'hours_until_departure']].copy()
@@ -906,7 +922,7 @@ def preprocess_data_simple(df_input, is_train=False):
         df_drop_nodes['high_price_duration_hours'] = prev_dur.loc[drop_mask].astype(float).to_numpy()
         df_drop_nodes['high_price_change_percent'] = prev_pct.loc[drop_mask].astype(float).round(4).to_numpy()
         df_drop_nodes['high_price_change_amount'] = prev_amo.loc[drop_mask].astype(float).round(2).to_numpy()
-
+        df_drop_nodes['high_price_seats_remaining_change_amount'] = prev_seats_amo.loc[drop_mask].astype(float).round(1).to_numpy()
         df_drop_nodes = df_drop_nodes.reset_index(drop=True)
 
         flight_info_cols = [
@@ -922,17 +938,20 @@ def preprocess_data_simple(df_input, is_train=False):
         df_drop_nodes = df_drop_nodes.merge(df_gid_info, on='gid', how='left')
 
         drop_info_cols = ['drop_hours_until_departure', 'drop_price_change_percent', 'drop_price_change_amount',
-                          'high_price_duration_hours', 'high_price_change_percent', 'high_price_change_amount'
+                          'high_price_duration_hours', 'high_price_change_percent', 'high_price_change_amount',
+                          'high_price_seats_remaining_change_amount',
         ]
         # 按顺序排列 去掉gid
         df_drop_nodes = df_drop_nodes[flight_info_cols + drop_info_cols]
+        df_drop_nodes = df_drop_nodes[df_drop_nodes['drop_price_change_percent'] <= -0.01]   # 太低的降幅不计
 
         # 对于没有先升后降的gid进行分析
         gids_with_drop = df_target.loc[drop_mask, 'gid'].unique()
         df_no_drop = df_target[~df_target['gid'].isin(gids_with_drop)].copy()
 
         keep_info_cols = [
-            'keep_hours_until_departure', 'keep_price_change_percent', 'keep_price_change_amount', 'keep_price_duration_hours'
+            'keep_hours_until_departure', 'keep_price_change_percent', 'keep_price_change_amount', 
+            'keep_price_duration_hours', 'keep_seats_remaining_change_amount',
         ]
         
         if df_no_drop.empty:
@@ -954,7 +973,8 @@ def preprocess_data_simple(df_input, is_train=False):
             )
 
             df_keep_nodes = df_keep_row[
-                ['gid', 'hours_until_departure', 'price_change_percent', 'price_change_amount', 'price_duration_hours']
+                ['gid', 'hours_until_departure', 'price_change_percent', 'price_change_amount', 
+                 'price_duration_hours', 'seats_remaining_change_amount']
             ].copy()
             df_keep_nodes.rename(
                 columns={
@@ -962,6 +982,7 @@ def preprocess_data_simple(df_input, is_train=False):
                     'price_change_percent': 'keep_price_change_percent',
                     'price_change_amount': 'keep_price_change_amount',
                     'price_duration_hours': 'keep_price_duration_hours',
+                    'seats_remaining_change_amount': 'keep_seats_remaining_change_amount',
                 },
                 inplace=True,
             )