Adaptive hybrid spatial hypergraph convolution module with data embedding optimization for stock ranking prediction

Abstract

Spatio-temporal data mining have various applications in the domains of finance, transportation, and sociology. Predicting stock rankings is a typical case that presents certain challenges. These challenges include: (1) The inability of existing graph learning methods, Regardless of how comprehensive their prior knowledge used for constructing graph relationships are, to generate significant improvements due to their lack of adaptive capturing of highdimensional data structures. (2) In time series forecasting. Stationarizing the data can more effectively capture data trends, but this operation may lead to the loss of important non-stationary factor information. (3) Spatio-temporal data mining models typically integrate time, space, and graph learning modules. Different modules with different functionalities often require different training environments. Many models rely solely on end-to-end optimization through the loss function. This leads to insufficient driving force for downstream modules, gradient environment disruptions, training blockages, and other issues. To address these challenges, we propose the Adaptive Hybrid Spatial Hypergraph Convolution Network (AHS-HGCN). Specifically, multiple multi-functional attention mechanisms are introduced to model the main task and provide suitable training environments for downstream modules. Among them, the HSCA module is capable of outputting hybrid spatial hyperedge weights and rewriting upstream outputs, maximizing the efficiency of model training. We evaluate the framework on three large-scale real stock datasets (NASDAQ, NYSE, and TSE). Compared to the baseline models, it achieved a minimum improvement of 27.22 % and a maximum improvement of 30.89 %.