Use categorical dtype for categorical features

The categorical 'actiontype', 'result', 'bodypart' and 'bodypart_detailed'
features now have dtype "pd.Categorical". This allows machine learning
frameworks such as xgboost to automatically recognize these features as being
categorical and handle them accordingly.

The features are now strings (e.g., "pass") instead of integer ids. Also, the
column name for each of these features was renamed as follows:
type_id -> actiontype
result_id -> result
bodypart_id -> bodypart

socceraction/vaep/features.py

@@ -6,7 +6,7 @@
 import pandas as pd  # type: ignore
 from pandera.typing import DataFrame
 
-import socceraction.spadl.config as spadlconfig
+import socceraction.spadl.config as spadlcfg
 from socceraction.atomic.spadl import AtomicSPADLSchema
 from socceraction.spadl.schema import SPADLSchema
 
@@ -117,9 +117,9 @@ def play_left_to_right(gamestates: GameStates, home_team_id: int) -> GameStates:
     away_idx = a0.team_id != home_team_id
     for actions in gamestates:
         for col in ['start_x', 'end_x']:
-            actions.loc[away_idx, col] = spadlconfig.field_length - actions[away_idx][col].values
+            actions.loc[away_idx, col] = spadlcfg.field_length - actions[away_idx][col].values
         for col in ['start_y', 'end_y']:
-            actions.loc[away_idx, col] = spadlconfig.field_width - actions[away_idx][col].values
+            actions.loc[away_idx, col] = spadlcfg.field_width - actions[away_idx][col].values
     return gamestates
 
 
@@ -169,7 +169,13 @@ def actiontype(actions: Actions) -> Features:
     Features
         The 'type_id' of each action.
     """
-    return actions[['type_id']]
+    X = pd.DataFrame(index=actions.index)
+    X["actiontype"] = pd.Categorical(
+        actions["type_id"].replace(spadlcfg.actiontypes_df().type_name.to_dict()),
+        categories=spadlcfg.actiontypes,
+        ordered=False,
+    )
+    return X
 
 
 @simple
@@ -187,8 +193,8 @@ def actiontype_onehot(actions: SPADLActions) -> Features:
         A one-hot encoding of each action's type.
     """
     X = {}
-    for type_id, type_name in enumerate(spadlconfig.actiontypes):
-        col = 'type_' + type_name
+    for type_id, type_name in enumerate(spadlcfg.actiontypes):
+        col = 'actiontype_' + type_name
         X[col] = actions['type_id'] == type_id
     return pd.DataFrame(X, index=actions.index)
 
@@ -207,7 +213,13 @@ def result(actions: SPADLActions) -> Features:
     Features
         The 'result_id' of each action.
     """
-    return actions[['result_id']]
+    X = pd.DataFrame(index=actions.index)
+    X["result"] = pd.Categorical(
+        actions["result_id"].replace(spadlcfg.results_df().result_name.to_dict()),
+        categories=spadlcfg.actiontypes,
+        ordered=False,
+    )
+    return X
 
 
 @simple
@@ -225,7 +237,7 @@ def result_onehot(actions: SPADLActions) -> Features:
         The one-hot encoding of each action's result.
     """
     X = {}
-    for result_id, result_name in enumerate(spadlconfig.results):
+    for result_id, result_name in enumerate(spadlcfg.results):
         col = 'result_' + result_name
         X[col] = actions['result_id'] == result_id
     return pd.DataFrame(X, index=actions.index)
@@ -275,10 +287,18 @@ def bodypart(actions: Actions) -> Features:
     bodypart_detailed :
         An alternative version that splits between the left and right foot.
     """
-    foot_id = spadlconfig.bodyparts.index("foot")
-    left_foot_id = spadlconfig.bodyparts.index("foot_left")
-    right_foot_id = spadlconfig.bodyparts.index("foot_right")
-    return actions[['bodypart_id']].replace([left_foot_id, right_foot_id], foot_id)
+    X = pd.DataFrame(index=actions.index)
+    foot_id = spadlcfg.bodyparts.index("foot")
+    left_foot_id = spadlcfg.bodyparts.index("foot_left")
+    right_foot_id = spadlcfg.bodyparts.index("foot_right")
+    X["bodypart"] = pd.Categorical(
+        actions["bodypart_id"]
+        .replace([left_foot_id, right_foot_id], foot_id)
+        .replace(spadlcfg.bodyparts_df().bodypart_name.to_dict()),
+        categories=["foot", "head", "other", "head/other"],
+        ordered=False,
+    )
+    return X
 
 
 @simple
@@ -303,7 +323,13 @@ def bodypart_detailed(actions: Actions) -> Features:
     bodypart :
         An alternative version that does not split between the left and right foot.
     """
-    return actions[['bodypart_id']]
+    X = pd.DataFrame(index=actions.index)
+    X["bodypart"] = pd.Categorical(
+        actions["bodypart_id"].replace(spadlcfg.bodyparts_df().bodypart_name.to_dict()),
+        categories=spadlcfg.bodyparts,
+        ordered=False,
+    )
+    return X
 
 
 @simple
@@ -328,19 +354,19 @@ def bodypart_onehot(actions: Actions) -> Features:
         An alternative version that splits between the left and right foot.
     """
     X = {}
-    for bodypart_id, bodypart_name in enumerate(spadlconfig.bodyparts):
+    for bodypart_id, bodypart_name in enumerate(spadlcfg.bodyparts):
         if bodypart_name in ('foot_left', 'foot_right'):
             continue
         col = 'bodypart_' + bodypart_name
         if bodypart_name == 'foot':
-            foot_id = spadlconfig.bodyparts.index("foot")
-            left_foot_id = spadlconfig.bodyparts.index("foot_left")
-            right_foot_id = spadlconfig.bodyparts.index("foot_right")
+            foot_id = spadlcfg.bodyparts.index("foot")
+            left_foot_id = spadlcfg.bodyparts.index("foot_left")
+            right_foot_id = spadlcfg.bodyparts.index("foot_right")
             X[col] = actions['bodypart_id'].isin([foot_id, left_foot_id, right_foot_id])
         elif bodypart_name == 'head/other':
-            head_id = spadlconfig.bodyparts.index("head")
-            other_id = spadlconfig.bodyparts.index("other")
-            head_other_id = spadlconfig.bodyparts.index("head/other")
+            head_id = spadlcfg.bodyparts.index("head")
+            other_id = spadlcfg.bodyparts.index("other")
+            head_other_id = spadlcfg.bodyparts.index("head/other")
             X[col] = actions['bodypart_id'].isin([head_id, other_id, head_other_id])
         else:
             X[col] = actions['bodypart_id'] == bodypart_id
@@ -370,17 +396,17 @@ def bodypart_detailed_onehot(actions: Actions) -> Features:
         An alternative version that does not split between the left and right foot.
     """
     X = {}
-    for bodypart_id, bodypart_name in enumerate(spadlconfig.bodyparts):
+    for bodypart_id, bodypart_name in enumerate(spadlcfg.bodyparts):
         col = 'bodypart_' + bodypart_name
         if bodypart_name == 'foot':
-            foot_id = spadlconfig.bodyparts.index("foot")
-            left_foot_id = spadlconfig.bodyparts.index("foot_left")
-            right_foot_id = spadlconfig.bodyparts.index("foot_right")
+            foot_id = spadlcfg.bodyparts.index("foot")
+            left_foot_id = spadlcfg.bodyparts.index("foot_left")
+            right_foot_id = spadlcfg.bodyparts.index("foot_right")
             X[col] = actions['bodypart_id'].isin([foot_id, left_foot_id, right_foot_id])
         elif bodypart_name == 'head/other':
-            head_id = spadlconfig.bodyparts.index("head")
-            other_id = spadlconfig.bodyparts.index("other")
-            head_other_id = spadlconfig.bodyparts.index("head/other")
+            head_id = spadlcfg.bodyparts.index("head")
+            other_id = spadlcfg.bodyparts.index("other")
+            head_other_id = spadlcfg.bodyparts.index("head/other")
             X[col] = actions['bodypart_id'].isin([head_id, other_id, head_other_id])
         else:
             X[col] = actions['bodypart_id'] == bodypart_id
@@ -450,8 +476,8 @@ def endlocation(actions: SPADLActions) -> Features:
     return actions[['end_x', 'end_y']]
 
 
-_goal_x: float = spadlconfig.field_length
-_goal_y: float = spadlconfig.field_width / 2
+_goal_x: float = spadlcfg.field_length
+_goal_y: float = spadlcfg.field_width / 2
 
 
 @simple
@@ -622,10 +648,10 @@ def goalscore(gamestates: GameStates) -> Features:
     actions = gamestates[0]
     teamA = actions['team_id'].values[0]
     goals = actions['type_name'].str.contains('shot') & (
-        actions['result_id'] == spadlconfig.results.index('success')
+        actions['result_id'] == spadlcfg.results.index('success')
     )
     owngoals = actions['type_name'].str.contains('shot') & (
-        actions['result_id'] == spadlconfig.results.index('owngoal')
+        actions['result_id'] == spadlcfg.results.index('owngoal')
     )
     teamisA = actions['team_id'] == teamA
     teamisB = ~teamisA

tests/vaep/test_features.py

@@ -40,13 +40,27 @@ def test_same_index(spadl_actions: DataFrame[SPADLSchema]) -> None:
         tm.assert_index_equal(features.index, spadl_actions.index)
 
 
+def test_actiontype(spadl_actions: DataFrame[SPADLSchema]) -> None:
+    gamestates = fs.gamestates(spadl_actions)
+    ltr_gamestates = fs.play_left_to_right(gamestates, 782)
+    out = fs.actiontype(ltr_gamestates)
+    assert out.shape == (len(spadl_actions), 3)
+
+
 def test_actiontype_onehot(spadl_actions: DataFrame[SPADLSchema]) -> None:
     gamestates = fs.gamestates(spadl_actions)
     ltr_gamestates = fs.play_left_to_right(gamestates, 782)
     out = fs.actiontype_onehot(ltr_gamestates)
     assert out.shape == (len(spadl_actions), len(spadl.config.actiontypes) * 3)
 
 
+def test_result(spadl_actions: DataFrame[SPADLSchema]) -> None:
+    gamestates = fs.gamestates(spadl_actions)
+    ltr_gamestates = fs.play_left_to_right(gamestates, 782)
+    out = fs.result(ltr_gamestates)
+    assert out.shape == (len(spadl_actions), 3)
+
+
 def test_result_onehot(spadl_actions: DataFrame[SPADLSchema]) -> None:
     gamestates = fs.gamestates(spadl_actions)
     ltr_gamestates = fs.play_left_to_right(gamestates, 782)
@@ -64,6 +78,13 @@ def test_actiontype_result_onehot(spadl_actions: DataFrame[SPADLSchema]) -> None
     )
 
 
+def test_bodypart(spadl_actions: DataFrame[SPADLSchema]) -> None:
+    gamestates = fs.gamestates(spadl_actions)
+    ltr_gamestates = fs.play_left_to_right(gamestates, 782)
+    out = fs.bodypart(ltr_gamestates)
+    assert out.shape == (len(spadl_actions), 3)
+
+
 def test_bodypart_onehot(spadl_actions: DataFrame[SPADLSchema]) -> None:
     gamestates = fs.gamestates(spadl_actions)
     ltr_gamestates = fs.play_left_to_right(gamestates, 782)