Robust Meta-Graph Learning: Exploring Multilevel Feature Interrelationships To Enhance Multimodal Time Series Prediction

The booming digital economy has provided unprecedented opportunities for enhancing prediction performance with multimodal time series data in a broad range of business analytics scenarios. In this study, we explore multilevel, asynchronous feature interrelationships in multimodal time series data and disentangle the intricate interrelationships with a design artifact named RMGL (for robust meta-graph learning). The design of RMGL is rooted in graph convolutional networks. The core of RMGL is a meta-graph composed of three hierarchically interconnected graphs, representing feature-, modality-, and timestep-wise interrelationships, respectively. The interconnections across the graphs allow feature representations to propagate simultaneously through a graph attention mechanism, thereby quantifying multilevel feature interrelationships with graph structures synchronously and efficiently. Furthermore, capitalizing on the low-pass nature of graph convolutional filters, RMGL introduces a weight regularization scheme to effectively learn the meta-graph for prediction. RMGL outperformed state-of-the-art alternatives in our empirical evaluation involving both regression (financial risk prediction) and classification (sentiment analysis, humor detection, and intent recognition) tasks. Ablation experiments, exploratory analyses, and economic value analysis indicate the utility of every design component in RMGL. History: Accepted by Ram Ramesh, Area Editor for Data Science and Machine Learning. Funding: G. Chen was supported by the National Natural Science Foundation of China [Grants 72522010, 72394371, 72301239, and 72550002]. S. Xiao was supported by the National Natural Science Foundation of China [Grants 72671226, 72301194, 72472058, and 72495133]. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2024.0933 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2024.0933 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ .