Research map / Time series / Long-term forecasting
Temporal Dependencies and Spatio-Temporal Patterns of Time Series Models
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.
A doctoral-thesis proposal summarizing three phases of interpretability work on deep time series forecasting models: attention-based spatio-temporal pattern analysis, sensitivity-based feature interaction analysis, and a planned windowed temporal saliency rescaling method.
Problem
Interpreting time series forecasting models is hard because of temporal dependencies between time steps and features whose importance changes over time; prior saliency methods often ignore temporal order or only use the last time step.
Approach
Proposes windowed temporal saliency rescaling, which computes the importance of each input time step separately relative to a prediction horizon and rescales feature importance across all look-back positions instead of assuming only the last time step matters.
Key results
- TFT outperformed four other deep learning models on all five regression metrics for predicting 15 days of COVID-19 cases from 13 days of input across 3,142 US counties.
- Sensitivity ranks of age-group features from the TFT model were reported to agree with CDC age-group case data, indicating high accuracy of the interpretation method.
- Self-attention weights from TFT were found to correlate with raw case data patterns and reporting artifacts and pandemic dynamics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Temporal Fusion Transformer (TFT) instead of the other four deep learning models | COVID-19 case forecasting · 3,142 US counties; predict the next 15 days from the previous 13 days | regression metrics | Helped |
| Compare sensitivity-based age-group rankings with CDC-reported age-group cases | TFT COVID-19 infection forecasting · US county-level COVID-19 data | ranking accuracy | Helped |
Limitations
- Windowed temporal saliency rescaling is proposed but not yet implemented or evaluated in this paper.
- Phase I relies on model-specific self-attention weights, which capture higher-level encoded patterns rather than direct input contributions.
- Evaluation of interpretation accuracy in Phase II is only via comparison to CDC age-group case rankings.
Experimental setup
Datasets: US county-level COVID-19 case data CDC COVID-19 age-group case reports
Metrics: five regression metrics
Baselines: Temporal Fusion Transformer
Builds on: Temporal Fusion Transformer temporal saliency methods
Closest papers by meaning
- SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.87)
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.86)
- TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation (ICML 2025 · similarity 0.86)
- Addressing Spatial-Temporal Heterogeneity: General Mixed Time Series Analysis via Latent Continuity Recovery and Alignment (NeurIPS 2024 · similarity 0.86)
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.86)
- Can LLMs Understand Time Series Anomalies? (ICLR 2025 · 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.