HybridGCN: An Integrative Model for Scalable Recommender Systems with Knowledge Graph and Graph Neural Networks

Graph Neural Networks (GNNs) have emerged as a state-of-the-art approach in building modern Recommender Systems (RS). By leveraging the complex relationships among items, users, and their attributes, which can be represented as a Knowledge Graph (KG), these models can explore implicit semantic sub-structures within graphs, thereby enhancing the learning of user and item representations. In this paper, we propose an end-to-end architectural framework for developing recommendation models based on GNNs and KGs, namely HybridGCN. Our proposed methodologies aim to address three main challenges: (1) making graph-based RS scalable on large-scale datasets, (2) constructing domain-specific KGs from unstructured data sources, and (3) tackling the issue of incomplete knowledge in constructed KGs. To achieve these goals, we design a multistage integrated procedure, ranging from user segmentation and LLM-supported KG construction process to interconnectedly propagating between the KG and the Interaction Graph (IG). Our experimental results on a telecom e-commerce domain dataset demonstrate that our approach not only makes existing GNN-based recommender baselines feasible on large-scale data but also achieves comparative performance with the HybridGCN core.