通过XAI驱动的数据归约实现时间序列分类的规模化扩展
文章背景与核心概要
尽管面向时间序列的可解释性人工智能(XAI)算法近年来取得了显著进展,但其在为下游任务提供可衡量的性能和计算收益方面的应用仍未得到充分探索。本文介绍了一种名为 drXAI 的新颖方法,该方法重新利用了 XAI 归因方法,以实现时间序列分类(TSC)中的高效数据归约。
现代最先进的模型(如 Transformer)由于对序列长度呈二次方复杂度、对通道数呈线性复杂度,在扩展性方面面临严峻挑战,这使得处理海量数据集在计算上代价高昂。drXAI 通过利用快速、GPU加速的分类器(Hydra)生成局部归因,将其聚合为全局特征重要性得分,并利用自动“肘部截断”(elbow-cut)启发式方法来选择显著特征,从而解决了这一难题。实验表明,drXAI 在真实数据集上实现了 80% 到 90% 的数据归约,同时保持了准确性,并使内存密集型模型(如 ConvTran)能够扩展到以前无法处理的大规模数据集上。
概述与总结
- arXiv ID: 2607.15774 [cs.LG]
- 作者: Davide Italo Serramazza, Thach Le Nguyen, Georgiana Ifrim
- 提交时间: 2026年7月17日(2026年8月13日最后修订)
- 状态: 已被 ECML-PKDD 2026 的 AALTD 研讨会接受
- 许可证: CC BY 4.0

总结
While Explainable AI (XAI) for time series has advanced significantly, its capacity to provide measurable performance and computational gains for downstream tasks remains largely unexplored. This paper introduces drXAI, a novel methodology that repurposes XAI attribution methods for effective data reduction in Time Series Classification (TSC).
尽管面向时间序列的可解释性人工智能(XAI)取得了显著进展,但其为下游任务提供可衡量的性能和计算收益的能力在很大程度上仍未得到探索。本文介绍了 drXAI,这是一种新颖的方法,它重新利用了 XAI 归因方法,用于时间序列分类(TSC)中的有效数据归约。
Modern state-of-the-art models like Transformers struggle with scalability due to quadratic complexity relative to sequence length and linear complexity relative channels, rendering massive datasets computationally prohibitive. drXAI solves this by leveraging a fast, GPU-accelerated classifier (Hydra) to generate local attributions, aggregating them into global feature importance scores, and utilizing an automated elbow-cut heuristic to select salient features. Experiments demonstrate that drXAI achieves 80% to 90% data reduction on real-world datasets while preserving accuracy and enabling memory-intensive models (such as ConvTran) to scale to previously inaccessible datasets.
由于相对于序列长度呈二次方复杂度、相对于通道数呈线性复杂度,Transformer 等现代最先进模型在扩展性方面举步维艰,这使得海量数据集在计算上令人望而却步。drXAI 通过利用快速的、GPU加速的分类器(Hydra)生成局部归因,将其聚合为全局特征重要性得分,并利用自动的肘部截断启发式算法来选择显著特征,从而解决了这个问题。实验证明,drXAI 在真实数据集上实现了 80% 到 90% 的数据归约,同时保持了准确性,并使内存密集型模型(如 ConvTran)能够扩展到以前无法访问的数据集。
摘要
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).
面向时间序列的可解释性人工智能(XAI)在算法上取得了显著增长,但其在为下游任务提供可衡量的性能收益方面的效用仍未得到充分探索。本文通过引入 drXAI 填补了这一空白,这是一种新颖的方法,它重新利用了 XAI 归因方法,以实现时间序列分类(TSC)中的有效数据归约。
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.
drXAIaddresses 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.
现代 TSC 的核心挑战是可扩展性;最先进的模型(如 Transformer)表现出与序列长度相关的二次方复杂度和与通道数相关的线性复杂度。这使得它们在处理海量数据集时计算成本过高。drXAI 通过使用快速的、GPU 加速的分类器(Hydra)生成局部归因来解决这个问题。我们将这些归聚合为全局特征重要性得分,并采用自动的肘部截断启发式算法来选择最显著的特征,而无需手动设置阈值。
We evaluate our approach on both synthetic and real-world univariate and multivariate datasets. On synthetic benchmarks,
drXAIsuccessfully recovers ground-truth features where traditional baselines fail. On real-world data,drXAIachieves between 80% and 90% data reduction while maintaining classification accuracy comparable to models trained on the full dataset. Most importantly, we show thatdrXAIallows 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.
我们在合成以及真实的单变量和多变量数据集上评估了我们的方法。在合成基准测试中,drXAI 成功恢复了传统基准失败的真实基础特征(ground-truth features)。在真实世界数据上,drXAI 实现了 80% 到 90% 的数据归约,同时保持了与在完整数据集上训练的模型相当的分类准确率。最重要的是,我们表明 drXAI 允许像 ConvTran 这样的资源密集型模型扩展到由于内存限制而以前无法访问的数据集。我们的结果展示了使用 XAI 的好处——不仅是为了可解释性,而且作为时间序列分析中特征选择和可扩展性的强大工具。我们所有的代码和数据都是公开可用的。
附加元数据
- 主要学科: 机器学习 (
cs.LG), 人工智能 (cs.AI) - DOI: 10.48550/arXiv.2607.15774
- 全文链接:
- 查看 PDF
- HTML 版本(实验性)
- TeX 源码