Improving Local Explainability By Learning Causal Graphs From Data

TMLR Paper6573 Authors

Abstract:

Causal Shapley values take into account causal relations among dependent features to adjust the contributions of each feature to a prediction. A limitation of this approach is that it can only leverage known causal relations.
In this work we combine the computation of causal Shapley values with causal discovery, i.e. learning causal graphs from data. In particular, we compute causal explanations across a set of candidate causal graphs learned from observational data, yielding a set of Shapley values that reflects the space of possible explanations consistent with the data. We propose two methods for estimating this list efficiently, drawing on the equivalences of the interventional distributions for a subset of the causal graphs. We evaluate our methods on synthetic and real-world data, showing that they provide explanations that are often closer to the true causal impacts compared to traditional Shapley value approaches that disregard causal relationships. Even when the discovered graph or MEC is imperfect, we on average observe improvements over marginal and conditional Shapley values.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Romain_Lopez1

Submission Number: 6573