Gradient-based Model Shortcut Detection for Time Series Classification

TL;DR

The paper addresses the problem of point-based shortcut learning in time series classification, showing that simple, spurious patterns can mislead deep models. It introduces the Shortcut Aggregate Gradient (SAG) score, a gradient-based method that aggregates input gradients from a pretrained model to detect class-level shortcuts without relying on test data or external attributes. Empirical results on UCR two-class datasets demonstrate high precision in identifying shortcut classes and datasets, with a clear demonstration on GunPoint and a SonyAIBO Robot Surface case study. The work provides a practical tool for diagnosing internal model biases in time series and offers a pathway toward more robust, reliable models in critical domains.

Abstract

Deep learning models have attracted lots of research attention in time series classification (TSC) task in the past two decades. Recently, deep neural networks (DNN) have surpassed classical distance-based methods and achieved state-of-the-art performance. Despite their promising performance, deep neural networks (DNNs) have been shown to rely on spurious correlations present in the training data, which can hinder generalization. For instance, a model might incorrectly associate the presence of grass with the label ``cat" if the training set have majority of cats lying in grassy backgrounds. However, the shortcut behavior of DNNs in time series remain under-explored. Most existing shortcut work are relying on external attributes such as gender, patients group, instead of focus on the internal bias behavior in time series models. In this paper, we take the first step to investigate and establish point-based shortcut learning behavior in deep learning time series classification. We further propose a simple detection method based on other class to detect shortcut occurs without relying on test data or clean training classes. We test our proposed method in UCR time series datasets.

Figures (8)

• Figure 1: Training and testing accuracies of ResNethe2016deep models on the GunPointUCRArchive dataset (top) and on the modified GunPoint dataset with a point shortcut feature added to all class 1 training samples (bottom) after 100 epochs.
• Figure 2: Resulting loss curves for ResNet18 model on Coffee, WormsTwoClass, and Yoga. Left: original datasets. Right: modified datasets with point shortcut features added to training only.
• Figure 3: Visualizations for datasets that are modified to include a point shortcut.
• Figure 4: SAG scores for regular dataset (left) and shortcut dataset (right) showing the effectiveness of parameter $\epsilon$ for filtering shortcut-affected samples.
• Figure 5: Visualizing the point-shortcut score on ResNet18 trained on (a) original d versus (b) point-shortcut injected to GunPoint data UCRArchive.