Cross-modal Transfer Learning from Tabular to Time Series Data via Semantic and Dynamic Temporal Graph Modeling

Abstract Accurate time series modeling plays a critical role in health, financial, and industrial applications. However, real-world deployment faces two key issues: (1) scarcity of clean training data, and (2) valuable information is locked in structured tabular formats, and cannot be imported into time series models. These limitations significantly constrain prediction performance in time series tasks. Cross-modal transfer learning (CMTL) provides a possible solution. Existing CMTL approaches succeed with image, text, and audio data by leveraging their inherent local correlations and hierarchical structures. However, tabular data presents unique difficulties. Specifically, the discrete, heterogeneous nature of tabular features, and lack of explicit temporal relationships in columns, prevent direct geometric alignment with time series waveforms. In fact, although it has no temporal nature, tabular data still contains rich domain knowledge that can be used by time series tasks. In this paper, we propose a cross-modal transfer learning (CMTL) model from tabular to time series data, called TSgraph. TSgraph first extracts the domain topology and feature interaction paradigm from tabular data, and constructs it as a graph structure. Then, the time series graph of each time step is dynamically updated in combination with these graphs, and finally get the classification results. We validate the performance of TSgraph on four real-world datasets from different domains and compare the results with state-of-the-art baseline models. Experiments show that our model achieves the best performance on multiple evaluation metrics.