Reproduce “TimeMKG: Knowledge-Infused Causal Reasoning for Multivariate Time Series Modeling” 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
Yifei Sun, Junming Liu, Yirong Chen, Xuefeng Yan, Ding Wang
TimeMKG is a multimodal causal reasoning framework for multivariate time series modeling that integrates variable-level textual semantics and explicit causal knowledge, structured in a Knowledge Graph via LLMs, with numerical time series data. It achieves superior performance in long-term forecasting, short-term forecasting, and classification tasks, demonstrating average MSE improvements of up to 15% compared to LLM-backboned methods and 71.0% accuracy in classification.
Key insights
- Integrating explicit, structured causal knowledge via an LLM-generated Knowledge Graph significantly enhances the predictive performance, interpretability, and generalization capabilities of multivariate time series models.
- LLMs are more effectively utilized for interpreting textual semantics and constructing structured knowledge (an LLM-powered approach) than directly processing numerical time series data (an LLM-backboned approach).
- Effective fusion of textual semantics and numerical observations at the variable level through cross-modality attention provides a deeper understanding of complex multivariate relationships compared to channel-independent or purely numerical models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.