CAGE: Causality-Aware Shapley Value for Global Explanations

TL;DR

The paper addresses the limitation of Shapley-based global explanations that ignore feature causality by introducing CAGE, a causality-aware global Shapley framework. It develops a post-interventional sampling scheme tied to a causal chain-graph representation, defines global causal Shapley values, proves four causal soundness properties, and provides an approximation algorithm. Through synthetic and ADNI data experiments, it demonstrates more faithful and intuitive explanations than traditional global methods, while acknowledging practical challenges such as obtaining and using reliable causal graphs. The work advances explainable AI by integrating causal reasoning into global explanations and outlines directions for learning causal structures under uncertainty to broaden applicability.

Abstract

As Artificial Intelligence (AI) is having more influence on our everyday lives, it becomes important that AI-based decisions are transparent and explainable. As a consequence, the field of eXplainable AI (or XAI) has become popular in recent years. One way to explain AI models is to elucidate the predictive importance of the input features for the AI model in general, also referred to as global explanations. Inspired by cooperative game theory, Shapley values offer a convenient way for quantifying the feature importance as explanations. However many methods based on Shapley values are built on the assumption of feature independence and often overlook causal relations of the features which could impact their importance for the ML model. Inspired by studies of explanations at the local level, we propose CAGE (Causally-Aware Shapley Values for Global Explanations). In particular, we introduce a novel sampling procedure for out-coalition features that respects the causal relations of the input features. We derive a practical approach that incorporates causal knowledge into global explanation and offers the possibility to interpret the predictive feature importance considering their causal relation. We evaluate our method on synthetic data and real-world data. The explanations from our approach suggest that they are not only more intuitive but also more faithful compared to previous global explanation methods.

Figures (4)

• Figure 1: Causal chain graph that shows the partial causal ordering of the Alzheimer dataset ADNIjack2008alzheimer used later in the experiments. The chain graph consists of three components $\tau_1$, $\tau_2$, and $\tau_3$ that have a causal ordering. In the components, the complete causal relationships of the variables are not known. Variables in $\tau_1$ and $\tau_3$ are assumed to have common confounders (green) and $\tau_2$ is assumed to have causal interaction (yellow). The target variable is marked in red.
• Figure 2: Results and data-generating causal structures for our experiments. The first row (Figures \ref{['fig:graphs_direct']}, \ref{['fig:graphs_markov']}, \ref{['fig:graphs_mix']}) show the true causal structure (left) of the data-generating SCMs and the corresponding causal chain graphs we use for the explanation (right). The second row (Figrues \ref{['fig:results_reg_direct']}, \ref{['fig:results_reg_markov']}, \ref{['fig:results_reg_mix']}, blue) shows the importance values determined for the linear regression models that were trained and evaluated on the causal structures above. The third row (Figures \ref{['fig:results_mlp_direct']}, \ref{['fig:results_mlp_markov']}, \ref{['fig:results_mlp_mix']}, green) shows the same information, but for the MLP models. The solid bars show values coming from SAGE, and the striped bars show values of our causally-aware global explanation method.
• Figure 3: Gold Standard Graph from shen2020challenges. The gold standard graph shows the causal relations between the seven features and the binary target variable DX. Blue nodes are biomarkers and white nodes are personal information about patients. From that, we derive the causal chain graph in Figure (\ref{['fig:causalchaingraph']}).
• Figure 4: Importance values of the ADNI data experiment. The left plot shows the feature importances for the MLP model and the right plot the feature importance of Random Forest. For each plot, the left set of bars shows the importance determined by SAGE and the right bars show the importances for our causality-aware global explanation method.

Theorems & Definitions (2)

• theorem thmcountertheorem
• proof