Reproduce “L-Drive: Beyond a Single Mapping-Latent Context Drives Time Series Forecasting” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fan Zhang, Shijun Chen, Hua Wang
L-Drive introduces a change-aware framework for multivariate time series forecasting that mitigates "response lag" in predictions during dynamic regime changes by explicitly modeling a Latent-Context of evolving patterns. The model consistently outperforms existing methods across various benchmarks, achieving up to 49.3% lower MSE on the ETTh2 dataset compared to DLinear, and demonstrates improved adaptability and computational efficiency.
Key insights
- Explicitly modeling dynamic changes through a latent context (L-Context) enables faster adaptation to regime switches and reduces prediction lag compared to direct mapping approaches.
- Modulating increment representations with an adaptively filtered change signal allows the model to promptly update its basis for future changes, preventing error accumulation during transitions.
- Incorporating learnable relative positional basis functions within patches improves the capture of local structural patterns, making the model more robust to diverse and evolving time series dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.