Research map / Time series / Foundation models for time series
Learning Soft Sparse Shapes for Efficient Time-Series Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SoftShape, which replaces hard shapelet sparsification with soft-weighted shape embeddings plus an MoE intra-shape learner and a shared Inception-based inter-shape learner for efficient, interpretable time series classification.
Problem
Shapelet-based time series classifiers discard candidate subsequences in a hard manner and ignore differing contributions of shapelets, while identifying shapelets is computationally expensive.
Approach
Soft shape sparsification weights shape embeddings by attention-derived contribution scores and fuses low-scored shapes into a single shape, then a learnable MoE router activates class-specific experts for intra-shape patterns while a shared Inception expert models inter-shape temporal dependencies over the shapes treated as a sequence.
Key results
- SoftShape achieves 0.9334 average test accuracy and 2.72 average rank across 128 UCR datasets, winning on 53 datasets, versus 0.9205 for TSLANet and 0.8972 for MR-H.
- Ablation on 128 UCR datasets: removing soft sparsification drops average accuracy to 0.9123, and removing both intra and inter learning drops it to 0.8696.
- Sparse ratio analysis on 18 datasets shows 50% sparsification leaves accuracy statistically unchanged from no sparsification (p>0.05), while 90% sparsity significantly degrades it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add soft shape sparsification that fuses lower-scored shapes (ablation) | SoftShape · UCR 128 | Avg. Acc: 0.9123 -> 0.9334 | Helped |
| add class-specific intra-shape experts (ablation) | SoftShape · UCR 128 | Avg. Acc: 0.9245 -> 0.9334 | Helped |
| add shared-expert inter-shape learning (ablation) | SoftShape · UCR 128 | Avg. Acc: 0.9022 -> 0.9334 | Helped |
| add intra- and inter-shape learning (ablation) | SoftShape · UCR 128 | Avg. Acc: 0.8696 -> 0.9334 | Helped |
| use a 1D CNN instead of a linear shape embedding layer (ablation) | SoftShape · UCR 128 | Avg. Acc: 0.9164 -> 0.9334 | Helped |
| increase sparsification from 0% to 10% (ablation) | SoftShape · 18 UCR datasets | Avg. Acc: 0.9461 -> 0.9469 | No clear effect |
| increase sparse ratio above 50%, to 70% or 90% (ablation) | SoftShape · 18 UCR datasets | Avg. Acc: 0.9461 -> 0.9323; 0.9461 -> 0.9261 | Hurt |
| activate more experts per shape (k=3 or k=4 rather than k=1 or k=2) (ablation) | SoftShape · 18 UCR datasets | Avg. Rank: 1.89 -> 2.78; 1.94 -> 2.39 | Hurt |
| set sliding step size to q=4 rather than q=1 (ablation) | SoftShape · 18 UCR datasets | Avg. Rank: 3.83 -> 1.78 | Helped |
| increase model depth from L=1 to L=2 (ablation) | SoftShape · 18 UCR datasets | Avg. Rank: 3.72 -> 2.61 | Helped |
| use multi-scale shape lengths (8, 16, 32) instead of validation-selected shape length (ablation) | SoftShape · 18 UCR datasets | Avg. Rank: 1.72 -> 1.44 | Mixed |
Surprising findings
- Sparsifying up to 50% of shape embeddings does not significantly reduce accuracy relative to the unsparsified model.
Limitations
- Only univariate time series are considered; multivariate relationships are not modeled.
- Runtime is not lower than TSLANet or MR-H in short-sequence and small-sample settings.
Experimental setup
Datasets: UCR time series archive (128 datasets)
Benchmarks: UCR 128 time series archive 18-dataset subset for LightTS and Shapeformer
Metrics: test accuracy average rank win count Wilcoxon signed-rank p-value runtime
Baselines: FCN T-Loss SelfTime TS-TCC TS2Vec TimesNet InceptionTime LightTS ShapeConv ModernTCN TSLANet TST
Builds on: shapelet transform mixture of experts InceptionTime patch-based time series models
Closest papers by meaning
- ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models (NeurIPS 2025 · similarity 0.89)
- Diffusion Language-Shapelets for Semi-supervised Time-Series Classification (AAAI 2024 · similarity 0.88)
- Abstracted Shapes as Tokens - A Generalizable and Interpretable Model for Time-series Classification (NeurIPS 2024 · similarity 0.88)
- Shape analysis for time series (NeurIPS 2024 · similarity 0.84)
- Soft Contrastive Learning for Time Series (ICLR 2024 · similarity 0.83)
- Learning to Embed Time Series Patches Independently (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.