Research map / Time series / Foundation models for time series
Towards Diverse Perspective Learning with Selection over Multiple Temporal Poolings
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 SoM-TP, a temporal pooling method that uses attention to dynamically select among global, static, and dynamic temporal poolings per batch, regularized by a perspective loss and a Diverse Perspective Learning Network, for time series classification.
Problem
Existing temporal pooling methods for CNN-based time series classification each use a fixed mechanism (global, uniform segmentation, or dynamic segmentation), so their performance varies across datasets and no single pooling dominates.
Approach
Combine GTP, STP and DTP outputs and select the best pooling per batch via attention, in a single classifier, using Multiple Choice Learning-inspired selection; a DPLN sub-network and a perspective loss (cross-entropy plus KL divergence) regularize the attention so that all pooling perspectives are reflected without iterative multi-model ensembling.
Key results
- ResNet SoM-TP MAX reaches 0.7773 accuracy on UCR and 0.6920 on UEA for FCN, outperforming the best single-pooling baselines (STP, DTP).
- SoM-TP achieves the lowest average rank (2.5) among compared pooling methods on both UCR and UEA, beating OS-CNN and other state-of-the-art TSC models on ROC AUC and PR AUC for UCR.
- On the ECG dataset SoM-TP MAX reaches 0.9566 accuracy, and on EEG 0.9510 accuracy, exceeding all single-perspective poolings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SoM-TP MAX instead of the strongest single temporal pooling | FCN · UCR (univariate) | ACC: 0.7462 -> 0.7556 | Helped |
| Use SoM-TP MAX instead of the strongest single temporal pooling | FCN · UEA (multivariate) | ACC: 0.6801 -> 0.6920 | Helped |
| Use SoM-TP MAX instead of the strongest single temporal pooling | ResNet · UCR (univariate) | ACC: 0.7456 -> 0.7773 | Helped |
| Use SoM-TP MAX instead of the strongest single temporal pooling | ResNet · UEA (multivariate) | ACC: 0.6717 -> 0.6769 | Helped |
| Use SoM-TP MAX instead of ROCKET | ResNet · UCR (univariate) | ACC: 0.7718 -> 0.7773 | Helped |
| Use SoM-TP MAX instead of OS-CNN | ResNet · UEA (multivariate) | ACC: 0.6808 -> 0.6920 | Helped |
| Add A0, phi0, DPLN, and L_attn to the variant with only phi0 (ablation) | SoM-TP · UCR (univariate) | ACC: 0.6966 -> 0.7503 | Helped |
| Add phi0, DPLN, and L_attn to the variant with only A0 (ablation) | SoM-TP · UCR (univariate) | ACC: 0.6963 -> 0.7503 | Helped |
| Add DPLN and L_attn to DPL Attention (ablation) | SoM-TP · UCR (univariate) | ACC: 0.6974 -> 0.7503 | Helped |
| Add A0 and L_attn to DPLN without A0 (ablation) | SoM-TP · UCR (univariate) | ACC: 0.7047 -> 0.7503 | Helped |
| Add L_attn to DPLN without L_attn (ablation) | SoM-TP · UCR (univariate) | ACC: 0.7399 -> 0.7503 | Helped |
| Use MAX rather than AVG selection operation (ablation) | SoM-TP · UCR/UEA repositories | ACC: 0.7773 vs 0.7659; 0.6920 vs 0.6764 | Mixed |
Surprising findings
- No single temporal pooling (GTP, STP, DTP) is dominant across datasets, and DTP is not always better than GTP or STP on the large ECG/EEG datasets.
Limitations
- Performance depends on the number of segments and the choice of pooling operations, which are tuned per dataset via greedy search.
- Interpretation via LRP is qualitative, with no ground truth for input attributions in TSC.
- Datasets with irregular lengths were excluded from the benchmark.
Experimental setup
Datasets: UCR archive UEA archive ECG EEG
Benchmarks: UCR UEA ECG EEG
Metrics: accuracy F1 macro ROC AUC PR AUC average rank
Baselines: GTP STP DTP ROCKET InceptionTime OS-CNN DSN MLSTM-FCN TCN Vanilla-Transformer TST ConvTran
Builds on: Multiple Choice Learning Dynamic Temporal Pooling soft-DTW FCN ResNet Layer-wise Relevance Propagation
Closest papers by meaning
- Pool Me Wisely: On the Effect of Pooling in Transformer-Based Models (NeurIPS 2025 · similarity 0.83)
- Simple and Efficient Heterogeneous Temporal Graph Neural Network (NeurIPS 2025 · similarity 0.81)
- Discrete Latent Perspective Learning for Segmentation and Detection (ICML 2024 · similarity 0.81)
- Single-Teacher View Augmentation: Boosting Knowledge Distillation via Angular Diversity (NeurIPS 2025 · similarity 0.81)
- Con4m: Context-aware Consistency Learning Framework for Segmented Time Series Classification (NeurIPS 2024 · similarity 0.81)
- TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning (ICML 2024 · similarity 0.80)
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.