[Paper Review] Causality Learning: A New Perspective for Interpretable Machine Learning

Guandong Xu, Tri Dung Duong+3 more|arXiv (Cornell University)|Jun 27, 2020

Explainable Artificial Intelligence (XAI)64 references41 citations

TL;DR

This paper surveys causal analysis foundations and causal approaches for interpretable machine learning, contrasting causality with traditional association-focused methods and outlining evaluation and open problems.

ABSTRACT

Recent years have witnessed the rapid growth of machine learning in a wide range of fields such as image recognition, text classification, credit scoring prediction, recommendation system, etc. In spite of their great performance in different sectors, researchers still concern about the mechanism under any machine learning (ML) techniques that are inherently black-box and becoming more complex to achieve higher accuracy. Therefore, interpreting machine learning model is currently a mainstream topic in the research community. However, the traditional interpretable machine learning focuses on the association instead of the causality. This paper provides an overview of causal analysis with the fundamental background and key concepts, and then summarizes most recent causal approaches for interpretable machine learning. The evaluation techniques for assessing method quality, and open problems in causal interpretability are also discussed in this paper.

Motivation & Objective

• Introduce the fundamental background and key concepts of causal analysis.
• Survey state-of-the-art causal approaches for interpretable machine learning.
• Discuss evaluation techniques for causal interpretability and outline open problems in the field.

Proposed method

• Present structural causal models and the potential outcome framework as bases for causal inference.
• Discuss treatment effect metrics including ITE, ATE, ATT, and CATE.
• Review model-agnostic and post-hoc causality methods for deep networks, including causal attribution and TCAV-style concepts.
• Summarize causal feature learning and counterfactual explanation techniques with objective functions and constraints.
• Describe visualization methods (e.g., PDP/ICE) for causal interpretation and outline evaluation frameworks.

Experimental results

Research questions

• RQ1 How can causal inference frameworks be integrated into interpretable ML to identify true causes of model decisions?
• RQ2 What are effective model-agnostic and post-hoc approaches to explainability grounded in causality?
• RQ3 How should we evaluate causal interpretability in the absence of ground-truth causal benchmarks?
• RQ4 What open problems remain in applying counterfactual explanations to classification, recommendation systems, and time-series data?

Key findings

• Causal approaches can identify causes and effects of model decisions beyond simple associations.
• There exist multiple frameworks for causal analysis, including structural causal models and potential outcomes, with established effect metrics (ITE, ATE, ATT, CATE).
• Several model-agnostic and post-hoc methods for causal interpretability are discussed, including causal feature learning, counterfactual explanations, and concept-based explanations.
• A variety of visualization and evaluation strategies are proposed for assessing causal interpretability, though ground-truth evaluation remains challenging.
• A range of toolboxes support causal analysis in ML, such as DoWhy, EconML, CausalNex, and TIGRAMITE.