Research map / Time series / Foundation models for time series
TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning
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.
Reformulates multivariate time series classification as weakly supervised multiple instance learning and proposes TimeMIL, a transformer-based time-aware MIL pooling with a learnable wavelet positional encoding.
Problem
Multivariate time series classification (MTSC) is usually treated as fully supervised learning, which ignores the sparsity and locality of discriminative time points and the temporal ordering among instances, while standard MIL pooling is permutation-invariant and ignores order.
Approach
Treat each time series as a bag and each time point as an instance, then use a tokenized transformer class token as a time-aware MIL pooling with multi-head self-attention for temporal correlation and a learnable wavelet positional encoding (WPE) for time-frequency ordering; multi-class is handled as one-vs-rest binary MIL, and attention weights give interpretability.
Key results
- Group 1 (26 UEA datasets): 77.4% average accuracy, best on 18 datasets and average rank 2.327 versus 4.058 for the second-best method (TodyNet).
- Group 2 (10 UEA datasets): 77.7% average accuracy, best on 5 datasets, average rank 2.85, outperforming ModernTCN by 3.5% in average accuracy.
- Ablation: time-aware pooling reaches 0.774 vs 0.739 for attention (ABMIL) pooling; WPE improves ABMIL and TimeMIL by 0.6% and 1% over no positional encoding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TimeMIL instead of the strongest listed baseline, TodyNet | MTSC models · Group 1 UEA datasets (26 datasets) | average accuracy: 0.726 -> 0.774 | Helped |
| Use TimeMIL instead of the strongest listed baseline, ModernTCN | MTSC models · Group 2 UEA datasets (10 datasets) | average accuracy: 0.742 -> 0.777 | Helped |
| Use TimeMIL instead of MOS-CNN | MTSC models · UEA datasets (30 datasets) | Accuracy: 0.78 -> 0.791 | Helped |
| Replace mean pooling with time-aware MIL pooling (ablation) | TimeMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.715 -> 0.774 | Helped |
| Replace max pooling with time-aware MIL pooling (ablation) | TimeMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.719 -> 0.774 | Helped |
| Replace attention pooling with time-aware MIL pooling (ablation) | TimeMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.739 -> 0.774 | Helped |
| Replace conjunctive pooling with time-aware MIL pooling (ablation) | TimeMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.746 -> 0.774 | Helped |
| Add sinusoidal positional encoding (ablation) | ABMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.739 -> 0.741 | No clear effect |
| Add wavelet positional encoding (ablation) | ABMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.739 -> 0.745 | Helped |
| Add sinusoidal positional encoding (ablation) | TimeMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.761 -> 0.763 | No clear effect |
| Add wavelet positional encoding (ablation) | TimeMIL · Group 1 UEA datasets (26 datasets) | Accuracy: 0.761 -> 0.774 | Helped |
Surprising findings
- Sinusoidal positional encoding gave only marginal gains (about 0.27%) over no positional encoding for both ABMIL and TimeMIL.
Limitations
- Does not model cross-channel information; the MIL formulation is applied per multivariate series without channel interactions.
- Some baseline results are N/A due to memory limits on long sequences, and baseline numbers are taken from prior papers.
Experimental setup
Datasets: UEA multivariate time series archive (26 equal-length datasets for Group 1; 10 datasets for Group 2) synthetic noisy pulse dataset
Benchmarks: UEA multivariate time series archive
Metrics: accuracy average accuracy average rank pairwise wins/draws/losses
Baselines: ED-1NN DTW-1NN-I DTW-1NN-D MLSTM-FCN ShapeNet WEASEL+MUSE TapNet OS-CNN MOS-CNN TodyNet ModernTCN PatchTST
Builds on: ABMIL Nyström self-attention conditional positional encoding OS-CNN
Closest papers by meaning
- TapNet: Multivariate Time Series Classification with Attentional Prototypical Network (AAAI 2020 · similarity 0.88)
- TAIL-MIL: Time-Aware and Instance-Learnable Multiple Instance Learning for Multivariate Time Series Anomaly Detection (AAAI 2025 · similarity 0.88)
- MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · similarity 0.88)
- Inherently Interpretable Time Series Classification via Multiple Instance Learning (ICLR 2024 · similarity 0.87)
- TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis (ICLR 2025 · similarity 0.86)
- T-Rep: Representation Learning for Time Series using Time-Embeddings (ICLR 2024 · similarity 0.85)
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.