[Paper Review] Preserving Causal Constraints in Counterfactual Explanations for Machine Learning Classifiers

Divyat Mahajan, Chenhao Tan+1 more|arXiv (Cornell University)|Dec 6, 2019

Explainable Artificial Intelligence (XAI)Computer Science16 references101 citations

TL;DR

The paper defines global and local feasibility for counterfactual explanations and introduces causal proximity regularizers and a VAE-based method to generate feasible counterfactuals under causal constraints, with user feedback when feasibility is unknown.

ABSTRACT

To construct interpretable explanations that are consistent with the original ML model, counterfactual examples---showing how the model's output changes with small perturbations to the input---have been proposed. This paper extends the work in counterfactual explanations by addressing the challenge of feasibility of such examples. For explanations of ML models in critical domains such as healthcare and finance, counterfactual examples are useful for an end-user only to the extent that perturbation of feature inputs is feasible in the real world. We formulate the problem of feasibility as preserving causal relationships among input features and present a method that uses (partial) structural causal models to generate actionable counterfactuals. When feasibility constraints cannot be easily expressed, we consider an alternative mechanism where people can label generated CF examples on feasibility: whether it is feasible to intervene and realize the candidate CF example from the original input. To learn from this labelled feasibility data, we propose a modified variational auto encoder loss for generating CF examples that optimizes for feasibility as people interact with its output. Our experiments on Bayesian networks and the widely used ''Adult-Income'' dataset show that our proposed methods can generate counterfactual explanations that better satisfy feasibility constraints than existing methods.. Code repository can be accessed here: extit{https://github.com/divyat09/cf-feasibility}

Motivation & Objective

• Define global and local feasibility for counterfactual explanations with respect to structural causal models.
• Introduce a causal proximity loss to preserve causal relationships during counterfactual generation.
• Propose a VAE-based method (Example-Based CF) that learns feasibility from user feedback when causal constraints are not explicit.
• Demonstrate that causal-aware methods produce more feasible counterfactuals on Adult-Income and Bayesian network datasets.
• Show that the proposed methods are faster at generating counterfactuals compared to some existing approaches.

Proposed method

• Formulate feasibility as constraints entailed by an underlying structural causal model (SCM).
• Propose DistCausal as a distance term that preserves causal relationships during CF generation.
• Provide unary and binary feasibility constraints as practical approximations when the full SCM is unknown.
• Introduce a model-based CF (with partial SCM) and a model-approx CF that incorporate feasibility.
• Develop Example-Based CF using a VAE to generate CFs and learn feasibility from user feedback through an oracle.
• Optimize a loss that combines validity, proximity, and feasibility, including a KL term in the VAE objective.

Experimental results

Research questions

• RQ1 How can counterfactual explanations respect causal constraints to remain feasible in the real world?
• RQ2 What is a practical proximal loss that enforces causal feasibility during CF generation?
• RQ3 Can user feedback be used to learn feasibility constraints when a full causal model is unavailable?
• RQ4 Do causal-aware CF methods yield more feasible counterfactuals and comparable validity to existing approaches?
• RQ5 How do the proposed methods perform on real and synthetic datasets (Adult, Simple-BN, Sangiovese) in terms of feasibility and efficiency?

Key findings

• Methods that incorporate causal feasibility outperform baseline methods in feasibility scores across datasets.
• All proposed feasibility-aware methods achieve perfect target-class validity on tested datasets.
• Model-based CF, Model-approx CF, and Example-Based CF generally yield higher Constraint Feasibility scores than CEM.
• Example-Based CF can learn feasibility from labeled examples and improve feasibility with more labels.
• The approaches show competitive or superior causal-edge scores on datasets with known causal structure.
• The proposed methods are faster at generating counterfactuals than some existing optimization-based baselines.