Presentation Information
[1Yin-B-13]Heterogeneous GNN Based on Content and Relational Edge Properties Considering Temporal Dynamics and Cross-Domain Transitions
〇Takumi Umehara1, Keishi Tajima1 (1. kyoto University)
Keywords:
Neural Network,Graph,Transfer Learning
This study aims to address composite problems involving temporal dynamics and cross-domain transitions by proposing HQRAN, a novel HGNN (Heterogeneous Graph Neural Network). The proposed model learns from a constructed heterogeneous graph where edges are categorized based on their properties into "Quantity Edges" containing numerical information and "Relational Edges" possessing structural information. Specifically, HQRAN employs a two-stage architecture to process these edges respectively: the first stage aggregates local interactions via Quantity Edges, while the second stage captures temporal and domain transitions via Relational Edges. To evaluate its effectiveness, we applied this approach to Major League Baseball (MLB) player performance prediction. Experiments predicting 2024 performance using data from 2016 to 2023 demonstrated that HQRAN achieved the highest accuracy compared to existing GNNs and traditional statistical methods. These findings conclude that categorizing learning mechanisms based on edge properties is effective for cross-domain time-series prediction tasks.