Feature-Importance Feature-Interactions (FIFI) graph: A graph-based Novel Visualization for Interpretable Machine Learning
Pawan Kumar, Manmohan Sharma · 2021 International Conference on Intelligent Technologies (CONIT) · 2021
A majority of the complex Machine Learning (ML) models lack in human interpretability. This lacking makes it difficult for human users to interpret why a particular prediction outcome has been made by an ML model. During recent years, a renewed interested has been observed among ML research community to confer human interpretability to ML models. Humans are inherently good at interpreting visualizations and have been using visualizations as a medium for explaining a phenomenon or process to others. Therefore, the output of most of the approaches towards conferring human interpretability to ML models has been usually a visualization. These visualizations aim to explain ML model behaviour to human users by highlighting the most contributing features. Studying interactions between features is equally important to facilitate knowledge discovery and providing new insights about the underlying physical phenomena. Existing visualizations mainly focusing on feature importance. This paper proposes a FIFI graph, a novel visualization for interpreting ML models. The intuitive idea is to plot the relative importance of features and relative strength of interactions between features into a single plot. The proposed visualization has been modelled as a network graph. It has been demonstrated using two datasets. Being a graph enables using existing research and software tools on network analysis for interpreting ML model behaviour. Also, a comparison of the FIFI graph with a knowledge graph has been made.