LRG Augments Features for Imbalanced Time Series Classification.

Imbalanced datasets pose a significant challenge in time series classification, often leading to poor performance on minority classes. While existing methods focus on adjusting training distributions or objectives, they often overlook a critical issue at the representation level: how reliable the learned feature space is around minority regions. This research identifies a "training-local geometry failure" where minority cases can end up in sparse or mixed feature-space neighborhoods, even if the overall class structure is preserved. To address this, the authors propose Local Reference Geometry (LRG), a lightweight, post-hoc feature augmentation module. LRG is applied between a fixed feature extractor and the classifier head. Using only training features, LRG assesses local exposure and class-mixture risk, then augments each feature with a standardized signed displacement from nearby training geometry and an LDA-projected residual summary. Empirical results on benchmarks show that LRG consistently improves performance for learned, pretrained, and fixed representations, particularly in high-risk minority regions, offering a novel solution to a persistent problem.