[Paper Review] Causality-Aware Local Interpretable Model-Agnostic Explanations

Martina Cinquini, Riccardo Guidotti|arXiv (Cornell University)|Dec 10, 2022

Topic Modeling4 citations

TL;DR

This paper proposes CALIME, a causality-aware extension of LIME that improves local model-agnostic explanations by generating synthetic data respecting causal relationships in the neighborhood of an instance. By replacing random perturbations with causal-aware data generation via gencda, CALIME achieves higher explanation fidelity and stability, especially as the number of features increases, though at the cost of increased computational time.

ABSTRACT

A main drawback of eXplainable Artificial Intelligence (XAI) approaches is the feature independence assumption, hindering the study of potential variable dependencies. This leads to approximating black box behaviors by analyzing the effects on randomly generated feature values that may rarely occur in the original samples. This paper addresses this issue by integrating causal knowledge in an XAI method to enhance transparency and enable users to assess the quality of the generated explanations. Specifically, we propose a novel extension to a widely used local and model-agnostic explainer, which encodes explicit causal relationships within the data surrounding the instance being explained. Extensive experiments show that our approach overcomes the original method in terms of faithfully replicating the black-box model's mechanism and the consistency and reliability of the generated explanations.

Motivation & Objective

• To address the limitation of existing XAI methods that assume feature independence and fail to model causal relationships in local explanations.
• To improve the fidelity and stability of local model-agnostic explanations by integrating causal structure into the synthetic data generation process.
• To develop a method that enhances trust in AI decisions by grounding explanations in realistic, causally informed data distributions.
• To demonstrate that causality-aware data generation leads to more plausible and robust explanations compared to random perturbation-based approaches.
• To lay the foundation for extending causality-aware explanation techniques beyond tabular data and continuous features.

Proposed method

• Replace LIME’s random synthetic data generation with gencda, a causal structure-aware synthetic data generator that respects the causal graph of the input features.
• Use causal discovery (via ncda) to infer the causal graph (DAG) from the training data before explanation.
• Generate synthetic instances in the local neighborhood of the instance to explain by sampling from the conditional distributions defined by the causal graph.
• Train a Lasso-regularized linear model on the causally generated synthetic data to produce feature importance explanations.
• Weight synthetic instances based on proximity to the original instance using a distance function, preserving local fidelity.
• Ensure the generated data respects the underlying causal dependencies, reducing the likelihood of generating implausible or unrealistic instances.

Experimental results

Research questions

• RQ1 Can integrating causal structure into synthetic data generation improve the fidelity of local model-agnostic explanations?
• RQ2 Does causality-aware data generation lead to more stable and consistent explanations across multiple runs?
• RQ3 How does the plausibility of synthetic instances generated by CALIME compare to those generated by LIME in terms of distributional realism?
• RQ4 What is the trade-off between explanation quality and computational efficiency when incorporating causal structure?
• RQ5 To what extent can causality-aware explanations improve user trust in AI decisions?

Key findings

• CALIME outperforms LIME in explanation plausibility across all metrics (AMD, AOD, ASM, ADM), with lower scores indicating higher plausibility, especially for datasets like statlog and wine-red.
• For fidelity, CALIME achieves higher average fidelity than LIME on all datasets except wdbc, where LIME slightly outperforms CALIME (0.44 vs. 0.41 for k=30).
• CALIME demonstrates superior stability for k > 4, with lower LLE (Local LIME Error) values, indicating more consistent explanations under varying perturbations.
• The neighborhood generated by CALIME is more compact and closer to the original instance than LIME’s, as shown by lower AMD and AOD scores, indicating higher local density and realism.
• Despite its advantages, CALIME is significantly slower than LIME, requiring one to two orders of magnitude more time due to causal graph inference and distribution learning.
• The method is currently limited to continuous features, as gencda and ncda do not support mixed-type (categorical/continuous) data, highlighting a key direction for future work.