Time series are generated continuously at massive scale by industrial equipment, wearables, power grids, and clinical monitors, yet annotation remains manual, costly, and expert‑dependent. Consequently, the bottleneck in large‑scale time series analytics is not data volume but label volume. Practitioners therefore face a concrete question: how many examples per class must be labeled for a classifier to become usable?
We address this question directly under a fixed, known label space, requiring the decision rule to be built from only K labeled examples per class. We propose the Dual‑Stream OSSE‑LSTM, an episodic metric‑learning framework that couples an Omni‑Scale CNN with Squeeze‑and‑Excitation recalibration for multi‑scale motif extraction—eliminating per‑dataset kernel tuning—with a Bidirectional LSTM for global temporal context. The two streams are independently normalized and fused into a prototype‑oriented embedding space.
Because decisions made from few labels must also be explainable, we introduce Counterfactual Integrated Gradients (C‑IG). Instead of attributing an isolated classifier logit, C‑IG attributes the prototype margin between the target class and opposing classes, and reuses the resulting attribution maps as soft masks for test‑time prototype refinement without updating the encoder.
Experiments on 19 univariate UCR datasets show that OSSE‑LSTM achieves the highest average accuracy and per‑dataset win count at every support size. Its accuracy stays within a narrow 0.36‑point band (96.36%‑96.72%) across the range. Even its weakest configuration surpasses the best result achieved by any baseline at any K (93.99%).
Review: The approach delivers strong performance under label‑scarce conditions while providing interpretability, offering a practical solution for large‑scale time series classification.