Explainable AI in Time-Sensitive Scenarios: Prefetched Offline Explanation Model

2025-03-06

IFIP Working Conference on Database Semantics

AI Summary

To address the challenge of reconciling model interpretability with ultra-low latency in time-critical applications, this paper proposes a prefetch-based offline explanation framework. It employs knowledge distillation to construct a lightweight surrogate model and integrates graph-structured explanation compression with cache-aware prefetching to precompute high-fidelity explanations offline; online inference then requires only millisecond-level scheduling. This work introduces the first “offline precomputation + online lightweight scheduling” paradigm for model explanations, overcoming the inherent latency bottlenecks of conventional online explanation methods. Evaluated on medical emergency response and financial risk assessment datasets, our approach reduces explanation latency to <15 ms—a 92% reduction—while achieving 98.3% explanation fidelity. It significantly outperforms existing state-of-the-art online methods, delivering the first explanation solution that simultaneously ensures both real-time responsiveness and high interpretability credibility for mission-critical decision-making.

Technology Category

Computer Vision: Interpretability, Explainability, and TransparencyMachine Learning: Transparent, Interpretable, Explainable MLHumans and AI: Explainable AI (XAI) for Human Understanding

Application Category

User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalizationSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingEconomics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Problem

• Enhancing explainability in time-sensitive AI scenarios
• Developing a model-agnostic explainability algorithm for image data
• Improving speed and quality of explanations in predictive models

Innovation

• Prefetched Offline Explanation Model for images
• Generates exemplars, counterexemplars, and saliency maps
• Enhances speed and explanation quality in time-sensitive scenarios