Research map / Time series / Foundation models for time series
Diffusion Language-Shapelets for Semi-supervised 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 DiffShape, a semi-supervised time-series classifier that learns shapelets via a real-subsequence-conditioned diffusion model and aligns them with T5 text embeddings of class descriptions through contrastive learning.
Problem
Semi-supervised time-series classification with limited labels, where learned shapelets are poorly discriminative and often unlike any real subsequence, hurting interpretability.
Approach
Uses the most similar real subsequences as a condition for a diffusion model that generates shapelets from unlabeled data, and adds contrastive language-shapelet alignment with natural-language class descriptions built from labels and pseudo-labels.
Key results
- On 106 UCR datasets at 10% labeling, DiffShape achieves average rank 2.92 with 51 wins, versus 3.24 and 25 wins for the previous best TS-TFC; at 40% it reaches rank 2.62 with 64 wins.
- On 12 UCR datasets with 10% labels, DiffShape ranks first (avg rank 1.00, 12 wins) against shapelet baselines ST, LTS, FSS and ADSN.
- Ablation on 12 datasets shows removing the diffusion mechanism raises avg rank from 1.00 to 3.08, and removing both mechanisms gives 5.17.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiffShape instead of TS-TFC | semi-supervised time-series classification · 106 UCR datasets, 10% labeled | average rank: 3.24 -> 2.92 | Helped |
| use DiffShape instead of TS-TFC | semi-supervised time-series classification · 106 UCR datasets, 20% labeled | average rank: 3.00 -> 2.86 | Helped |
| use DiffShape instead of TS-TFC | semi-supervised time-series classification · 106 UCR datasets, 40% labeled | average rank: 2.88 -> 2.62 | Helped |
| use DiffShape instead of ST | shapelet-based semi-supervised time-series classification · 12 UCR datasets, 10% labeled | average rank: 2.92 -> 1.00 | Helped |
| use DiffShape instead of SemiTime | semi-supervised time-series classification without unlabeled data · 12 UCR datasets, 2 labels per class | average rank: 3.50 -> 1.42 | Helped |
| use DiffShape instead of TS-TFC | semi-supervised time-series classification without unlabeled data · 12 UCR datasets, 5 labels per class | average rank: 2.58 -> 1.67 | Helped |
| use DiffShape instead of MTFC | semi-supervised time-series classification without unlabeled data · 12 UCR datasets, 10 labels per class | average rank: 2.83 -> 1.75 | Helped |
| add self-supervised diffusion learning (ablation) | DiffShape · 12 UCR datasets, 10% labeled | average rank: 3.08 -> 1.00 | Helped |
| generate shapelets instead of replacing them with most similar real subsequences (ablation) | DiffShape · 12 UCR datasets, 10% labeled | average rank: 2.83 -> 1.00 | Helped |
| condition diffusion on most similar real subsequences rather than random real subsequences (ablation) | DiffShape · 12 UCR datasets, 10% labeled | average rank: 3.17 -> 1.00 | Helped |
| add contrastive language-shapelets learning (ablation) | DiffShape · 12 UCR datasets, 10% labeled | average rank: 3.67 -> 1.00 | Helped |
| add both self-supervised diffusion and contrastive language-shapelets learning (ablation) | DiffShape · 12 UCR datasets, 10% labeled | average rank: 5.17 -> 1.00 | Helped |
Surprising findings
- MTL and MTFC, which use unsupervised time prediction losses, fail to improve over the supervised baseline and rank poorly.
Limitations
- Only univariate time series are studied; multivariate extension is future work.
- Pseudo-labels may be incorrect, which the method mitigates only by selecting high-confidence predictions.
Experimental setup
Datasets: UCR time series archive (106 datasets)
Benchmarks: UCR time series archive
Metrics: test classification accuracy average rank win count Wilcoxon signed-rank p-value
Baselines: Supervised Pseudo-Label Temporal Ensembling LPDeepSSL MTL TS-TCC SemiTime SSSTC MTFC TS-TFC ST LTS
Builds on: shapelet learning (LTS) DDIM SemiTime TS-TFC SimCLR
Closest papers by meaning
- ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models (NeurIPS 2025 · similarity 0.90)
- Abstracted Shapes as Tokens - A Generalizable and Interpretable Model for Time-series Classification (NeurIPS 2024 · similarity 0.89)
- Learning Soft Sparse Shapes for Efficient Time-Series Classification (ICML 2025 · similarity 0.88)
- TimeDART: A Diffusion Autoregressive Transformer for Self-Supervised Time Series Representation (ICML 2025 · similarity 0.88)
- Language Pre-training Guided Masking Representation Learning for Time Series Classification (AAAI 2025 · similarity 0.86)
- Retrieval-Augmented Diffusion Models for Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
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.