Research map / Time series / Foundation models for time series
MF-CLR: Multi-Frequency Contrastive Learning Representation for Time Series
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 MF-CLR, a hierarchical self-supervised contrastive representation method for multi-frequency time series, using a Dual Twister augmentation and cross-frequency consistency losses.
Problem
Learning self-supervised representations from unlabeled time series whose channels are sampled at different frequencies (e.g., daily, monthly, quarterly financial data), which existing uni-frequency contrastive methods handle poorly.
Approach
A hierarchical block-wise encoder that processes frequency groups from highest to lowest, combining instance-wise contrast on a Dual Twister augmented view (temporal and value warping that preserves frequency structure) with a cross-frequency contrast that pulls representations together according to lower-frequency distances, and concatenating lower-frequency projections into the next block.
Key results
- On 30 UEA classification datasets, MF-CLR reaches 0.6487 accuracy, beating the best self-supervised baseline T-Loss (0.6044) and the supervised PatchTST (0.6540 is higher, so MF-CLR is second among supervised only on some metrics).
- On anomaly detection, MF-CLR achieves F1 0.6798 vs. 0.6576 for TS2Vec (second best self-supervised).
- In the two real-world case studies (stock price and net income MAPE), MF-CLR ranks top two in 7 of 8 subtasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Dual Twister augmentation (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: MSE change + 0.080 when removed | Helped |
| Add cross-frequency contrast (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: MSE change + 0.107 when removed | Helped |
| Add low-frequency projector (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: MSE change + 0.095 when removed | Helped |
| Add hierarchical contrast (ablation) | MF-CLR · ETTm2 hourly OT, forecasting, QH-D | MSE: MSE change + 0.011 when removed | Helped |
| Add hierarchical contrast (ablation) | MF-CLR · ETTm2 hourly OT, forecasting, Q-HD | MSE: MSE change + 0.056 when removed | Helped |
| Use TCN instead of MLP encoder (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: MLPs: + 0.077 | Helped |
| Use TCN instead of LSTM encoder (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: LSTM: + 0.081 | Helped |
| Use TCN instead of Transformer encoder (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: Transformer: + 0.014 | Helped |
| Add temporal contrastive loss (ablation) | MF-CLR · ETTm2 hourly OT, forecasting | MSE: + 0.102 | Hurt |
| Use MF-CLR rather than TS2Vec | time-series anomaly detection · SMD, SMAP, MSL, and SWaT | F1: 0.6576 -> 0.6798 | Helped |
| Use MF-CLR rather than PatchTST | classification · 30 UEA datasets | accuracy: 0.6540 -> 0.6487 | Hurt |
Surprising findings
- Adding temporal contrastive loss degrades forecasting performance substantially (MSE +0.102 on ETTm2), contrary to common temporal-contrast practice.
- MF-CLR underperforms supervised baselines on imputation for several datasets, such as Informer on ETTm1 and ETTm2.
Limitations
- Multi-frequency datasets were built by resampling public uni-frequency datasets, not native multi-frequency data.
- Does not lead on traffic imputation against self-supervised baselines and falls behind some supervised methods in imputation.
- Temporal consistency is intentionally excluded, which limits fine-grained temporal modeling.
Experimental setup
Datasets: UEA (30 datasets) SMD SMAP MSL SWaT traffic ETTm1 ETTm2 weather US and China stock price and net income data
Benchmarks: UEA classification SMD SMAP MSL SWaT traffic ETTm1 ETTm2 weather
Metrics: accuracy macro F1 MSE F1 score MAPE
Baselines: T-Loss TS-TCC TNC CPC TS2Vec CoST TF-C DeepAR TCN Informer Autoformer FEDformer
Builds on: TS2Vec TS-TCC TF-C CoST TNC
Closest papers by meaning
- Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning (AAAI 2025 · similarity 0.87)
- Self-Supervised Contrastive Learning for Long-term Forecasting (ICLR 2024 · similarity 0.86)
- TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting (ICLR 2025 · similarity 0.85)
- CFPT: Empowering Time Series Forecasting through Cross-Frequency Interaction and Periodic-Aware Timestamp Modeling (ICML 2025 · similarity 0.84)
- Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.84)
- Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · similarity 0.84)
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.