TimeMKG is a multimodal causal inference framework that leverages semantic information in variables to improve the performance and interpretability of time series data modeling. Unlike existing time series models, which treat variables as anonymous statistical signals, TimeMKG uses a large-scale language model to interpret the semantic information contained in variable names and data descriptions, constructing a structured multivariate knowledge graph that captures relationships among variables. It employs a bimodal encoder that separately models semantic prompts generated from knowledge graph triplets and statistical patterns in historical time series data. It then aligns and fuses these representations at the variable level to inject causal prior information into downstream tasks such as prediction and classification. Experiments on diverse datasets demonstrate that incorporating variable-level knowledge improves both predictive and generalization performance.
