A Static-Dynamic Hypergraph Neural Network Framework Based on Residual Learning for Stock Recommendation

Abstract

Stock ranking prediction is an effective method for achieving a high investment return and plays a crucial role in investment decisions. However, previous studies have overlooked the interconnections among stocks or have solely relied on predefined graphs for stock relationship information. The predefined graphs may not capture all possible relationships and may not be suitable for describing dynamic relationships. To address these issues, we propose a Static-Dynamic hypergraph neural network framework based on Residual Learning (SD-RL). Compared with traditional methods, SD-RL has the following advantages. (1) Stocks are no longer treated as isolated entities; instead, their static and dynamic relationship information is taken into account. (2) Leveraging the data-driven methodology, SD-RL autonomously learns both the static graph and dynamic hypergraph through dedicated graph learning and hypergraph learning modules, respectively. (3) By employing residual learning, various latent relationship information flows are mined, which enhances the stock embedding’s capacity to capture trends. Extensive experiments on the real data verify the effectiveness of our proposed methods.