Interpretable Machine Learning Based on Functional ANOVA Framework: Algorithms and Comparisons

In the early days of machine learning (ML), the emphasis was on developing complex algorithms to achieve best possible predictive performance. To understand and explain the model results, one had to rely on post hoc explainability techniques, which are known to have limitations. Recently, with the recognition in regulated industries that interpretability is also important, researchers are studying algorithms that compromise on small increases in predictive performance in favor of being more interpretable. While doing so, the ML community has rediscovered the use of low-order functional ANOVA (fANOVA) models that have been known in the statistical literature for some time. This paper starts with a description of challenges with post hoc explainability. This is followed by a brief review of the fANOVA framework with a focus on models with just main effects and second-order interactions (called generalized additive models with interactions or GAMI = GAM + Interactions). It then provides an overview of two recently developed GAMI techniques: Explainable Boosting Machines or EBM and GAMI-Net. The paper proposes a new algorithm that also uses trees, as in EBM, but does linear fits instead of piecewise constants within the partitions. We refer to this as GAMI-linear-tree (GAMI-Lin-T). There are many other differences, including the development of a new interaction filtering algorithm. The paper uses simulated and real datasets to compare the three fANOVA ML algorithms. The results show that GAMI-Lin-T and GAMI-Net have comparable performances, and both are generally better than EBM.