This application is a simple example of inspecting global feature importance for binary and multiclass classification with KNIME Software. The key of this example is the Global Feature Importance component verified and developed by the KNIME Team. In this example, the Wine quality data set is partitioned to training and test samples. Then, the black box model (Neural Network) is trained on the standardly pre-processed training data using the AutoML component. The Workflow Object capturing the pre-processing and the model is provided as an input for the Global Feature Importance component together with the test data. The component provides the global feature importance according to four techniques: three interpretable Global Surrogate Models (GLM, Decision Tree, and Random Forest) and Permutation Feature Importance.
Hi @ribizli
Not correct! The component is able to explain any classifier that takes as input data and spits out predictions with proper renaming of columns.
To feed your custom model to the component, you need to capture the workflow segment with the predictor node and any data preparation step via integrated deployment. To learn how to use Integrated Deployment read here:
It sounds a bit confusing at first because you want to explain a model, not to deploy it, but what integrated deployment truly does is capturing the model even just to export and save it or to flexibly reapply it on the fly with any model-agnostic technique. To read more about the multi purpose of Integrated Deployment read here:
A blog post is soon coming out to explain more in detail how to use the Global Feature Importance component.
@Lada do we have an example workflow where the component is applied to a model trained without the AutoML component? If not we should not make one!
Have a look at this example workflow on the KNIME Hub. The Global Feature Importance component is used for a custom model captured in the Workflow Object.
