[WIP] Debiasing BYOR

transformers/mli/debiasing_lfr.py

@@ -0,0 +1,100 @@
+from h2oaicore.transformer_utils import CustomTransformer
+
+import datatable as dt
+import numpy as np
+
+
+class LfrDebiasingTransformer(CustomTransformer):
+    _regression = False
+    _multiclass = False
+
+    _modules_needed_by_name = ['aif360']
+
+    _display_name = "LrfDebiasingTransformer"
+
+    @staticmethod
+    def get_default_properties():
+        return dict(
+            col_type="all",
+            min_cols="all",
+            max_cols="all",
+            relative_importance=1,
+        )
+
+    @staticmethod
+    def do_acceptance_test():
+        return False
+
+    def fit(self, X: dt.Frame, y: np.array = None):
+        from h2oaicore.systemutils import config
+        from aif360.datasets import BinaryLabelDataset
+        from aif360.algorithms.preprocessing.lfr import LFR
+
+        if y is not None:
+            if 'recipe_dict' in config:
+                config = config['recipe_dict']
+
+            # LFR supports only numerical columns
+            # But categoricals which are numeric are ok so setting col_type="all"
+            if any(unsupported in str(X.ltypes) for unsupported in ['str', 'obj']):
+                return
+
+            X_pd = X.to_pandas()
+            X = dt.Frame(X_pd.fillna(X_pd.mean()))
+
+            frame = dt.cbind(X, dt.Frame(y))
+            self.label_names = [frame.names[-1]]
+
+            self.privileged_groups = config['privileged_groups']
+            self.unprivileged_groups = config['unprivileged_groups']
+            self.favorable_label = float(config['favorable_label'])
+            self.unfavorable_label = float(config['unfavorable_label'])
+            self.protected_attribute_names = config['protected_attribute_names']
+
+            self.lfr = LFR(
+                unprivileged_groups=self.unprivileged_groups,
+                privileged_groups=self.privileged_groups,
+                verbose=0,
+            )
+
+            self.lfr.fit(
+                BinaryLabelDataset(
+                    df=frame.to_pandas(),
+                    favorable_label=self.favorable_label,
+                    unfavorable_label=self.unfavorable_label,
+                    label_names=self.label_names,
+                    protected_attribute_names=self.protected_attribute_names,
+                )
+            )
+            self.fitted = True
+
+    def fit_transform(self, X: dt.Frame, y: np.array = None):
+        self.fit(X, y)
+        return self.transform(X, y)
+
+    def transform(self, X: dt.Frame, y: np.array = None):
+        from aif360.datasets import BinaryLabelDataset
+        # Transformation should only occur during training when y is present
+        if self.fitted and (self.label_names in X.names or y is not None):
+            if self.label_names not in X.names:
+                X = dt.cbind(X, dt.Frame(y))
+
+            X_pd = X.to_pandas()
+            X = dt.Frame(X_pd.fillna(X_pd.mean()))
+            transformed_X: BinaryLabelDataset = self.lfr.transform(
+                BinaryLabelDataset(
+                    df=X.to_pandas(),
+                    favorable_label=self.favorable_label,
+                    unfavorable_label=self.unfavorable_label,
+                    label_names=self.label_names,
+                    protected_attribute_names=self.protected_attribute_names,
+                )
+            )
+
+            return dt.Frame(
+                transformed_X.features,
+                names=[name+"_lfr" for name in transformed_X.feature_names],
+            )
+        # For predictions no transformation is required
+        else:
+            return X