Beyond single perspective bias: Fusing personalized and common preferences for comprehensive personal preference learning.

In recommendation systems, Graph Convolutional Network (GCN)-based models are generally influenced by popular items. Over-emphasizing these items can lead to a single-perspective bias that overshadows the learning of the user's personalized preferences. Therefore, existing GCN-based models usually suppress information from popular items. However, as popular items with rich interactions contain the user's common preference information, such approaches may introduce another single-perspective bias that neglects the learning of the user's common preferences. Contrary to the prevailing assumption, we argue that personalized and common preferences are not mutually exclusive. Thus, we propose P&CGCN to collaboratively fuse them within a unified framework. This unified framework includes two parts: intra-layer aggregation and inter-layer combination. Specifically, in intra-layer aggregation, we design P&C degree to quantify the manifestation of personal preferences in each item, adaptively discerning whether it reflects personalized or common preferences without explicit separation. The P&C degree-based intra-layer aggregation guides context-aware integration of both preference aspects at each layer. In inter-layer combination, we design P&C depth to quantify the importance of each layer. The P&C depth-based inter-layer combination systematically prioritizes shallow-layer personalized preference signals while strategically leveraging deep-layer common preference signals. Comparative experiments on four real-world datasets demonstrate the performance and efficiency of P&CGCN. In particular, on sparse large datasets, the performance of P&CGCN has improved by around 20 % compared to LightGCN, with at least a 2x speedup in training efficiency.