Directly Optimizing Explanations for Desired Properties

2024-10-31

arXiv.org

AI Summary

Existing black-box model explanation methods heuristically encourage fidelity, conciseness, fairness, and other desirable properties—yet they lack rigorous guarantees and cannot controllably resolve multi-objective trade-offs. This paper introduces the first end-to-end differentiable explanation optimization framework, directly modeling target attributes as differentiable loss terms and enabling users to explicitly specify multi-attribute weights for customizable trade-off control. Our core contributions are: (1) a gradient-differentiable feature attribution module; (2) a weighted multi-objective joint optimization objective; and (3) hard constraints and dynamic adjustment mechanisms for attribute satisfaction rates. Extensive evaluation across multiple benchmarks demonstrates significant improvements in both attribute compliance rates and explanation stability. To our knowledge, this is the first work to enable controllable generation of explanation quality and task-adaptive customization.

Technology Category

• Humans and AI: Explainable AI (XAI) for Human Understanding
• Computer Vision: Interpretability, Explainability, and Transparency
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

When explaining black-box machine learning models, it's often important for explanations to have certain desirable properties. Most existing methods `encourage' desirable properties in their construction of explanations. In this work, we demonstrate that these forms of encouragement do not consistently create explanations with the properties that are supposedly being targeted. Moreover, they do not allow for any control over which properties are prioritized when different properties are at odds with each other. We propose to directly optimize explanations for desired properties. Our direct approach not only produces explanations with optimal properties more consistently but also empowers users to control trade-offs between different properties, allowing them to create explanations with exactly what is needed for a particular task.

Problem

• Inconsistent achievement of desired explanation properties
• Lack of control over property prioritization in explanations
• Need for optimizing explanations for specific property trade-offs

Innovation

• Directly optimize explanations for desired properties
• Enable user control over property trade-offs
• Ensure consistent optimal property attainment