def produce_shap_values(
self,
*,
inputs: container.DataFrame,
timeout: float = None,
iterations: int = None,
) -> CallResult[container.DataFrame]:
if self._needs_fit:
self.fit()
# don't want to produce SHAP predictions on train set because too computationally intensive
check_rows = min(self._input_hash.shape[0], inputs.shape[0])
if (pd.util.hash_pandas_object(
inputs.head(check_rows)) == self._input_hash.head(check_rows)
).all():
logger.info(
"Not producing SHAP interpretations on train set because of computational considerations"
)
return CallResult(container.DataFrame([]))
# drop any non-numeric columns
num_cols = inputs.shape[1]
inputs = inputs.select_dtypes(include="number")
col_diff = num_cols - inputs.shape[1]
if col_diff > 0:
logger.warn(f"Removed {col_diff} unencoded columns.")
explainer = shap.TreeExplainer(self._model._models[0].model)
max_size = self.hyperparams["shap_max_dataset_size"]
if inputs.shape[0] > max_size:
logger.warning(
f"There are more than {max_size} rows in dataset, sub-sampling ~{max_size} approximately representative rows "
+ "on which to produce interpretations")
df = self._shap_sub_sample(inputs)
shap_values = explainer.shap_values(df)
else:
shap_values = explainer.shap_values(pd.DataFrame(inputs))
if self._model.mode == "classification":
logger.info(
f"Returning interpretability values offset from most frequent class in dataset"
)
shap_values = shap_values[np.argmax(explainer.expected_value)]
output_df = container.DataFrame(shap_values, generate_metadata=True)
for i, col in enumerate(inputs.columns):
output_df.metadata = output_df.metadata.update_column(
i, {"name": col})
component_cols: Dict[str, List[int]] = {}
for c in range(0, len(output_df.columns)):
col_dict = dict(
inputs.metadata.query((metadata_base.ALL_ELEMENTS, c)))
if "source_column" in col_dict:
src = col_dict["source_column"]
if src not in component_cols:
component_cols[src] = []
component_cols[src].append(c)
# build the source column values and add them to the output
for s, cc in component_cols.items():
src_col = output_df.iloc[:, cc].apply(lambda x: sum(x), axis=1)
src_col_index = len(output_df.columns)
output_df.insert(src_col_index, s, src_col)
output_df.metadata = output_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, src_col_index),
"https://metadata.datadrivendiscovery.org/types/Attribute",
)
df_dict = dict(output_df.metadata.query(
(metadata_base.ALL_ELEMENTS, )))
df_dict_1 = dict(
output_df.metadata.query((metadata_base.ALL_ELEMENTS, )))
df_dict["dimension"] = df_dict_1
df_dict_1["name"] = "columns"
df_dict_1["semantic_types"] = (
"https://metadata.datadrivendiscovery.org/types/TabularColumn", )
df_dict_1["length"] = len(output_df.columns)
output_df.metadata = output_df.metadata.update(
(metadata_base.ALL_ELEMENTS, ), df_dict)
return CallResult(output_df)
def produce(
self,
*,
inputs: container.DataFrame,
timeout: float = None,
iterations: int = None,
) -> CallResult[container.DataFrame]:
logger.debug(f"Producing {__name__}")
# force a fit it hasn't yet been done
if self._needs_fit:
self.fit()
# drop any non-numeric columns
# drop all non-numeric columns
num_cols = inputs.shape[1]
inputs = inputs.select_dtypes(include="number")
col_diff = num_cols - inputs.shape[1]
if col_diff > 0:
logger.warn(
f"Removed {col_diff} unencoded columns from produce data.")
# create dataframe to hold the result
result = self._model.predict(inputs.values)
if len(self._target_cols) > 1:
result_df = container.DataFrame()
for i, c in enumerate(self._target_cols):
col = container.DataFrame({c: result[:, i]})
result_df = pd.concat([result_df, col], axis=1)
for c in range(result_df.shape[1]):
result_df.metadata = result_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, c), "http://schema.org/Float")
else:
result_df = container.DataFrame({self._target_cols[0]: result},
generate_metadata=True)
# if we mapped values earlier map them back.
if len(self._label_map) > 0:
# TODO label map will not work if there are multiple output columns.
result_df[self._target_cols[0]] = result_df[
self._target_cols[0]].map(self._label_map)
# mark the semantic types on the dataframe
for i, _ in enumerate(result_df.columns):
result_df.metadata = result_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, i),
"https://metadata.datadrivendiscovery.org/types/PredictedTarget",
)
if (self._model.mode == "classification"
and self.hyperparams["compute_confidences"]):
confidence = self._model.predict_proba(inputs.values)
if self._binary:
pos_column = (0 if self.hyperparams["pos_label"]
== self._label_map[0] else 1)
result_df.insert(result_df.shape[1], "confidence",
confidence[:, pos_column])
result_df.metadata = result_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, len(result_df.columns) - 1),
"http://schema.org/Float",
)
else:
# add confidence scores as some metrics require them.
confidence = pd.Series(confidence.tolist(), name="confidence")
result_df = pd.concat([result_df, confidence], axis=1)
confidences = [
item
for sublist in result_df["confidence"].values.tolist()
for item in sublist
]
labels = np.array(
list(self._label_map.values()) * len(result_df))
index = [
item for sublist in [[i] * len(np.unique(labels))
for i in result_df.index]
for item in sublist
]
result_df_temp = container.DataFrame()
result_df_temp["Class"] = labels
result_df_temp["confidence"] = confidences
result_df_temp.metadata = result_df.metadata
result_df_temp["index_temp"] = index
result_df_temp = result_df_temp.set_index("index_temp")
result_df = result_df_temp
result_df.metadata = result_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, len(result_df.columns) - 1),
"https://metadata.datadrivendiscovery.org/types/FloatVector",
)
result_df.metadata = result_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, len(result_df.columns) - 1),
"https://metadata.datadrivendiscovery.org/types/Score",
)
result_df.metadata = result_df.metadata.add_semantic_type(
(metadata_base.ALL_ELEMENTS, len(result_df.columns) - 1),
"https://metadata.datadrivendiscovery.org/types/PredictedTarget",
)
logger.debug(f"\n{result_df}")
return base.CallResult(result_df)
