以關聯圖卷積機制建構可解釋知識感知循環網絡於推薦系統

Abstract

本研究提出一套混合式圖推薦模型，結合了嵌入式方法與路徑式方法，以彌補兩者各自的限制。嵌入式方法擅長透過鄰近聚合捕捉語意擴散，但缺乏可解釋性；路徑式方法則能執行多跳推理，然而易受到路徑爆炸與訓練與推論分布不一致的影響。為整合兩者優勢，本研究採用關聯圖卷積網路（R-GCN），產生具語境化的實體與關係嵌入，並用以引導路徑選擇與序列建模。此外，設計一組自訂損失函數，透過全域關係率（Global Relationship Rate）平衡任務導向與語意擴散的學習需求。為解決路徑過載問題並對齊訓練與推論分布，本研究提出一套以嵌入為基礎的可信度評分機制作為路徑選擇模組。實驗結果顯示，在 MovieLens 與 Amazon Musical Instruments 資料集上，本模型於 precision@10 與 NDCG@10 指標皆穩定優於各項基線方法，且在弱勢使用者的推薦品質（CVAR@10）上也有顯著改善。此外，由於路徑序列直接參與預測過程，本模型天生具備可解釋性，有助於提升推薦透明度與使用者信任。

This study proposes a hybrid model for graph-based recommendation that integrates embedding-based and path-based approaches to address their respective limitations. While the former captures diffusive semantics via neighborhood aggregation but lacks interpretability, the latter enables multi-hop reasoning yet suffers from path explosion and train-inference distribution mismatch. To bridge these gaps, we incorporate a Relational Graph Convolutional Network (R-GCN) to generate contextualized entity and relation embeddings, which guide both path selection and sequential modeling. A custom loss function balances task-specific and global relationship learning through the Global Relationship Rate. To mitigate the path overload and align path distributions between training and inference, we design a path selection module based on embedding-informed credibility scores. Experiments on MovieLens and Amazon Musical Instruments datasets show that our model consistently outperforms baselines in precision@10 and NDCG@10, while also improving fairness for the worst-served users as measured by CVAR@10. Furthermore, as path sequences are used directly in prediction, the model inherently provides interpretable recommendations, enhancing transparency and trust.