Research map / Time series / Foundation models for time series
Explaining Time Series via Contrastive and Locally Sparse Perturbations
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 ContraLSP, a perturbation-based time series explainer that learns counterfactual perturbations via a triplet contrastive loss and sample-specific locally sparse gates with a learned temporal trend smoothing term.
Problem
Explaining multivariate time series models by identifying salient time-feature locations; existing perturbation methods produce out-of-distribution or informative perturbations, causing unfaithful explanations, especially with heterogeneous samples.
Approach
Learn a perturbation function whose outputs are counterfactual via inverse triplet contrastive loss (pulled toward negative-class samples, pushed from anchor's positives) to keep perturbations in-domain and uninformative, combined with sample-specific l0-relaxed stochastic sparse gates whose logits are modulated by a learned trend function for smooth binary-skewed masks.
Key results
- On Rare-Observation, ContraLSP reaches AUR 1.00 and mask entropy 0.32 (x10^-2), versus Extrmask AUR 0.76 and entropy 9.55.
- On Switch-Feature and State, ContraLSP is best on 7 of 8 metrics against strongest baselines, with AUR 0.80 and 0.81 respectively.
- On MIMIC-III (20% masked, average substitution), ContraLSP improves Acc to 0.980 and CE to 0.127 over Extrmask (0.982, 0.118).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add contrastive triplet loss (ablation) | ContraLSP · Switch-Feature | AUP: 0.97_{\pm0.01} -> 0.98_{\pm0.00} | Helped |
| add contrastive triplet loss (ablation) | ContraLSP · State | AUP: 0.88_{\pm0.03} -> 0.90_{\pm0.03} | Helped |
| add trend function (ablation) | ContraLSP · Switch-Feature | AUP: +0.06 | Helped |
| add trend function (ablation) | ContraLSP · State | AUP: +0.13 | Helped |
| use Euclidean distance instead of Manhattan distance in contrastive loss (ablation) | ContraLSP · Rare-Time | I_m: 19.51_{\pm0.30} -> 19.67_{\pm0.52} | No clear effect |
| use cosine similarity instead of Manhattan distance in contrastive loss (ablation) | ContraLSP · Rare-Time | I_m: 19.51_{\pm0.30} -> 18.41_{\pm0.64} | Hurt |
| increase regularization factors to alpha=1.0 and beta=2.0 (ablation) | ContraLSP · Switch-Feature | AUP: 0.53_{\pm0.05} -> 0.98_{\pm0.01} | Helped |
| use alpha=2.0 and beta=1.0 (ablation) | ContraLSP · State | AUP: 0.54_{\pm0.01} -> 0.90_{\pm0.02} | Helped |
| use alpha=0.005 and beta=0.01 (ablation) | ContraLSP · MIMIC-III mortality, average substitution | CE: 0.124_{\pm0.007} -> 0.127_{\pm0.007} | Mixed |
Limitations
- Label-free counterfactual generation may not yield sufficiently strong counterfactuals
- Sparsity hyperparameters (alpha, beta) must be tuned per dataset
- Evaluation relies on synthetic ground truth plus proxy metrics on MIMIC-III without true attributions
Experimental setup
Datasets: Rare-Time (synthetic) Rare-Observation (synthetic) Switch-Feature (synthetic) State (synthetic) MIMIC-III
Benchmarks: Rare-Time Rare-Observation Switch-Feature State MIMIC-III mortality
Metrics: AUP AUR information I_m mask entropy S_m Accuracy Cross-Entropy Sufficiency Comprehensiveness
Baselines: Feature Occlusion Augmented Feature Occlusion Integrated Gradients Shapley Value Sampling Dynamask Extrmask GradSHAP DeepLIFT LIME FIT RETAIN
Builds on: Extrmask Dynamask stochastic gates (STG) triplet loss preservation game
Closest papers by meaning
- Addressing Spatial-Temporal Heterogeneity: General Mixed Time Series Analysis via Latent Continuity Recovery and Alignment (NeurIPS 2024 · similarity 0.87)
- Optimal Information Retention for Time-Series Explanations (ICML 2025 · similarity 0.87)
- Explain Temporal Black-Box Models via Functional Decomposition (ICML 2024 · similarity 0.86)
- Causality-Aware Contrastive Learning for Robust Multivariate Time-Series Anomaly Detection (ICML 2025 · similarity 0.86)
- Empowering Time Series Analysis with Large-Scale Multimodal Pretraining (NeurIPS 2026 · similarity 0.86)
- CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series (ICML 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.