AI Insight
This paper introduces drXAI, a methodology that uses explainable AI attribution methods to reduce data dimensionality in time series classification tasks. The approach employs a fast classifier to generate local feature attributions, aggregates them into global importance scores, and uses an automated selection heuristic to identify the most relevant features. Testing on synthetic and real-world datasets demonstrates that drXAI achieves 80-90% data reduction while maintaining classification accuracy and enables complex models to process previously computationally prohibitive datasets.
Why it matters
This work addresses a critical scalability bottleneck in time series analysis, allowing resource-intensive models to handle massive datasets that were previously inaccessible due to memory constraints. The methodology demonstrates practical utility of explainable AI beyond interpretability, offering a systematic approach to feature selection that could accelerate applications in healthcare monitoring, industrial sensor analysis, and other domains with large-scale temporal data.
Understand the Science
arXiv:2607.15774v2 Announce Type: replace
Abstract: Explainable AI (XAI) for time series has seen significant algorithmic growth, but its utility in providing measurable performance gains for downstream tasks remains under-explored. This paper bridges this gap by introducing drXAI, a novel methodology that repurposes XAI attribution methods for effective data reduction in Time Series Classification (TSC). The core challenge in modern TSC is scalability; state-of-the-art models, such as Transformers, exhibit quadratic complexity relative to sequence length and linear complexity relative to the number of channels. This renders them computationally prohibitive for massive datasets. drXAI addresses this by using a fast, GPU-accelerated classifier (Hydra) to generate local attributions. We aggregate these into global feature importance scores and employ an automated elbow-cut heuristic to select the most salient features without requiring manual thresholds.
We evaluate our approach on both synthetic and real-world univariate and multivariate datasets. On synthetic benchmarks, drXAI successfully recovers ground-truth features where traditional baselines fail. On real-world data, drXAI achieves between 80% and 90% data reduction while maintaining classification accuracy comparable to models trained on the full dataset. Most importantly, we show that drXAI allows resource-intensive models like ConvTran to scale to datasets that were previously inaccessible due to memory constraints. Our results show the benefits of using XAI not just for interpretability, but as a robust tool for feature selection and scalability in time series analysis. All our code and data are openly available.
Source: Scaling Time Series Classification via XAI-Driven Data Reduction