Answer: how to apply shap Python package to DSS visual model

[MarkPundurs]

Here's the code I used for a visual KNN model:

# -*- coding: utf-8 -*-
import dataiku
import pandas as pd, numpy as np
from dataiku import pandasutils as pdu
import shap

# Read recipe inputs
model_1 = dataiku.Model("l10JO8hN")
# get DSS Predictor from model
pred_1 = model_1.get_predictor()
feat_ds = dataiku.Dataset("myDatasetName")
# get pandas dataframe from DSS dataset
feat_df = feat_ds.get_dataframe()
# get numpy array from dataframe
feat_arr = feat_df.values

# Compute recipe outputs from inputs:

# define function that takes a numpy array as its argument and returns the output of the DSS Predictor
def shap_predict(feat_arr):
    # rebuild dataframe from array
    shap_df = pd.DataFrame.from_records(feat_arr, columns=feat_df.columns)
    # pass dataframe to DSS Predictor and return output
    return pred_1.predict(shap_df)

# pass array-to-prediction function and feature array as arguments to shap.KernelExplainer()
explainer = shap.KernelExplainer(shap_predict, feat_arr)
shap_values = explainer.shap_values(feat_arr)
shap_df = pd.DataFrame.from_records(shap_values)

# Write recipe outputs
shap = dataiku.Dataset("shap_new")
shap.write_with_schema(shap_df)

[AdrienL]

Hi,
This is not fully supported for visual models, may work for some but not sure it will work for all kinds of models / preprocessings. Thanks for sharing the code that did work for you though.
Note that DSS 12.1 natively offers shapley values estimations and their export into a dataset.