Merge pull request #269 from microsoft/encode-nominals

Encoding categorical nominal to integers on discover DECI/Notears/PC backend

python/backend/backend/discover/algorithms/commons/base_runner.py

@@ -83,6 +83,16 @@ def _remove_rows_with_missing_values(self):
             self._number_of_rows - self._prepared_data.shape[0]
         )
 
+    def _transform_categorical_nominal_to_continuous(self):
+        # TODO: remove this once categorical values are properly handled by each algorithm
+        for name in self._prepared_data.columns:
+            if (
+                self._nature_by_variable[name]
+                == CausalVariableNature.CategoricalNominal
+            ):
+                logging.info(f"encoding categorical nominal column {name} to integers")
+                self._prepared_data[name] = pd.factorize(self._prepared_data[name])[0]
+
     def _prepare_data(self):
         self._prepared_data = pd.DataFrame.from_dict(self._dataset_data)
         self._remove_rows_with_missing_values()

python/backend/backend/discover/algorithms/deci.py

@@ -89,6 +89,7 @@ def __init__(self, p: DeciPayload, progress_callback: ProgressCallback = None):
         self._is_dag = None
 
     def _build_causica_dataset(self) -> Dataset:
+        self._transform_categorical_nominal_to_continuous()
         numpy_data = self._prepared_data.to_numpy()
         data_mask = np.ones(numpy_data.shape)
 

python/backend/backend/discover/algorithms/notears.py

@@ -20,6 +20,8 @@ def __init__(self, p: NotearsPayload, progress_callback: ProgressCallback = None
         super().__init__(p, progress_callback)
 
     def do_causal_discovery(self) -> CausalGraph:
+        self._transform_categorical_nominal_to_continuous()
+
         notears_graph = from_pandas(
             self._prepared_data,
             tabu_child_nodes=self._constraints.causes,

python/backend/backend/discover/algorithms/pc.py

@@ -23,6 +23,8 @@ def __init__(self, p: PCPayload, progress_callback: ProgressCallback = None):
         super().__init__(p, progress_callback)
 
     def do_causal_discovery(self) -> CausalGraph:
+        self._transform_categorical_nominal_to_continuous()
+
         n = PC(alpha=0.2)
         n.learn(self._prepared_data.to_numpy())
         graph_gc = networkx.DiGraph(n.causal_matrix)