XAI-Driven Data Reduction Boosts Time Series Classification Scalability

Modern Time Series Classification (TSC) models, especially Transformers, face significant scalability challenges due to their computational complexity with increasing sequence length and channels. A novel approach, drXAI, addresses this by leveraging Explainable AI (XAI) techniques for intelligent data reduction. This method uses a fast, GPU-accelerated classifier to generate local feature attributions, which are then aggregated into global importance scores. An automated "elbow-cut" heuristic selects the most relevant features, eliminating the need for manual thresholds. Evaluations on both synthetic and real-world datasets demonstrate drXAI's effectiveness. It successfully identifies ground-truth features in synthetic benchmarks and achieves 80-90% data reduction on real-world data while preserving classification accuracy. Crucially, drXAI enables powerful but resource-intensive models, like ConvTran, to process datasets that were previously too large due to memory constraints, highlighting XAI's utility beyond just interpretability.