Causal Dependence Plots for Interpretable Machine Learning

Abstract

Explaining artificial intelligence or machine learning models is an increasingly important problem. For humans to stay in the loop and control such systems, we must be able to understand how they interact with the world. This work proposes using known or assumed causal structure in the input variables to produce simple and practical explanations of supervised learning models. Our explanations-which we name Causal Dependence Plots or CDP-visualize how the model output depends on changes in a given predictor along with any consequent causal changes in other predictors. Since this causal dependence captures how humans often think about input-output dependence, CDPs can be powerful tools in the explainable AI or interpretable ML toolkit and contribute to applications including scientific machine learning and algorithmic fairness. CDP can also be used for model-agnostic or black-box explanations.Preliminary work under review.

The benefits and pitfalls of machine learning for biomarker discovery

“…However, with even minimal background knowledge (e.g. a partial ordering over variable classes), we can use XAI tools to quantify causal impact and disambiguate direct from indirect effects (Loftus et al 2023). Moreover, these methods can integrate with existing procedures to generate new hypotheses that can be tested using more traditional methods such as gene knockout experiments.…”