Research map / Time series / Foundation models for time series
Con4m: Context-aware Consistency Learning Framework for Segmented 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 Con4m, a consistency learning framework for segmented time series classification on multi-class varying-duration data, combining context-aware Gaussian-prior attention, coherent prediction constraints, and curriculum-based progressive label harmonization to handle inconsistent boundary labels.
Problem
Segmented time series classification where raw series contain multiple classes of varying duration (MVD); existing TSC models assume i.i.d. segments and ignore temporal dependencies, while inconsistent expert boundary labels destabilize training.
Approach
Introduces a Con-Attention layer mixing self-attention with a learnable per-segment Gaussian kernel prior for locality, a context-aware coherent prediction with neighbor-pair same-class discrimination and a Tanh-fitted monotonicity constraint, and a curriculum-based progressive label update that harmonizes boundary labels from easy core to hard transition segments.
Key results
- On the SEEG raw labels, Con4m reaches 72.00% F1 vs 67.64% for the next best (Scale-teaching).
- On Sleep with 0% disturbance Con4m achieves 68.02% F1, a 4.81-point gain over Scale-teaching (63.21%) per the table.
- Label substitution: harmonized labels improve PatchTST by 4.11% and TimesNet by 7.53% on SEEG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Con4m rather than the strongest listed baseline | Segmented TSC · HHAR 0% label disturbance | F1 score (%): 78.61 -> 80.29 | Helped |
| Use Con4m rather than the strongest listed baseline | Segmented TSC · Sleep 0% label disturbance | F1 score (%): 63.48 -> 68.02 | Helped |
| Use Con4m rather than the strongest listed baseline | Segmented TSC · SEEG raw labels | F1 score (%): 67.64 -> 72.00 | Helped |
| Add Tanh function fitting to the prediction behavior constraint (ablation) | Con4m · Sleep 20% label disturbance | F1 score (%): 65.28 -> 66.31 | Helped |
| Add Tanh function fitting to the prediction behavior constraint (ablation) | Con4m · Sleep 40% label disturbance | F1 score (%): 63.22 -> 64.31 | Helped |
| Add Tanh function fitting to the prediction behavior constraint (ablation) | Con4m · SEEG raw labels | F1 score (%): 70.63 -> 72.00 | Helped |
| Use progressive label updating with η (ablation) | Con4m · Sleep 20% label disturbance | F1 score (%): 59.97 -> 66.31 | Helped |
| Use progressive label updating with η (ablation) | Con4m · Sleep 40% label disturbance | F1 score (%): 57.98 -> 64.31 | Helped |
| Use progressive label updating with η (ablation) | Con4m · SEEG raw labels | F1 score (%): 66.04 -> 72.00 | Helped |
| Add the full method's modules rather than retaining only one module (ablation) | Con4m · Inconsistent scenarios (Sleep and SEEG) | F1 score (%): averaged 2.78% decrease | Helped |
Surprising findings
- Removing the label-update weight eta (no progressive label harmonization) causes the largest drop (about 9.2% F1), suggesting early use of model predictions as labels causes error accumulation.
- TAS models (MS-TCN2) are competitive on fNIRS but degrade sharply on HHAR, Sleep and SEEG with more ambiguous boundaries.
Limitations
- Only end-to-end supervised models are studied; large-scale models are not explored.
- The Tanh function-fitting module needs more basis function design for more diverse label behaviors.
- Private SEEG dataset cannot be released, and only one test group receives majority-vote labels.
Experimental setup
Datasets: fNIRS HHAR SleepEDF SEEG
Benchmarks: fNIRS HHAR SleepEDF SEEG
Metrics: F1 score accuracy C-score
Baselines: MS-TCN2 ASFormer DiffAct MiniRocket TimesNet PatchTST SIGUA UNICON Sel-CL SREA Scale-teaching
Builds on: Scale-teaching SREA MS-TCN2 TimesNet PatchTST
Closest papers by meaning
- Context-Alignment: Activating and Enhancing LLMs Capabilities in Time Series (ICLR 2025 · similarity 0.84)
- InteDisUX: Intepretation-Guided Discriminative User-Centric Explanation for Time Series (AAAI 2025 · similarity 0.84)
- FIC-TSC: Learning Time Series Classification with Fisher Information Constraint (ICML 2025 · similarity 0.84)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.83)
- Multi-Patch Prediction: Adapting Language Models for Time Series Representation Learning (ICML 2024 · similarity 0.83)
- TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series (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.