[Paper Review] Knowledge-aware Neural Collective Matrix Factorization for Cross-domain Recommendation

Li Zhang, Ge Yan+3 more|arXiv (Cornell University)|Jun 27, 2022

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes KG-NeuCMF, a knowledge-aware neural collective matrix factorization framework that integrates knowledge graphs (KGs) to enhance cross-domain recommendation by jointly learning domain-specific and domain-general item representations from KG-structured knowledge. By leveraging graph autoencoders and mutual information maximization between user-item interactions and KG embeddings, the model significantly outperforms state-of-the-art baselines, achieving up to 21% relative improvement in MAE on cold-start and sparse scenarios.

ABSTRACT

Cross-domain recommendation (CDR) can help customers find more satisfying items in different domains. Existing CDR models mainly use common users or mapping functions as bridges between domains but have very limited exploration in fully utilizing extra knowledge across domains. In this paper, we propose to incorporate the knowledge graph (KG) for CDR, which enables items in different domains to share knowledge. To this end, we first construct a new dataset AmazonKG4CDR from the Freebase KG and a subset (two domain pairs: movies-music, movie-book) of Amazon Review Data. This new dataset facilitates linking knowledge to bridge within- and cross-domain items for CDR. Then we propose a new framework, KG-aware Neural Collective Matrix Factorization (KG-NeuCMF), leveraging KG to enrich item representations. It first learns item embeddings by graph convolutional autoencoder to capture both domain-specific and domain-general knowledge from adjacent and higher-order neighbours in the KG. Then, we maximize the mutual information between item embeddings learned from the KG and user-item matrix to establish cross-domain relationships for better CDR. Finally, we conduct extensive experiments on the newly constructed dataset and demonstrate that our model significantly outperforms the best-performing baselines.

Motivation & Objective

• Address the limitations of existing cross-domain recommendation (CDR) models that rely solely on common users or mapping functions, which fail to exploit rich semantic relationships between items across domains.
• Overcome the challenge of negative transfer in bidirectional CDR by incorporating knowledge graphs (KGs) as a natural bridge to share domain-general and domain-specific knowledge.
• Develop a unified framework that aligns item representations learned from user-item interaction matrices with those derived from KGs to improve cross-domain recommendation performance.
• Construct a new public benchmark dataset, AmazonKG4CDR, to enable reproducible research in KG-aware cross-domain recommendation.
• Improve performance in cold-start and data-sparse scenarios by leveraging KG-based item representations that remain informative even when user interaction data is limited.

Proposed method

• Construct the AmazonKG4CDR dataset by integrating Freebase KG with two domain pairs (movies-music, movies-books) from Amazon Review Data, enabling cross-domain item linking via shared entities and relations.
• Employ a relational graph convolutional autoencoder (RGCN-based) to learn item embeddings from the KG, capturing both adjacent and higher-order neighborhood information across domains.
• Introduce a neural mutual information estimator to maximize correlation between item embeddings from the user-item interaction matrix and those from the KG, ensuring effective cross-domain alignment.
• Unify the learned KG-based item representations with user-item interaction embeddings in a neural collective matrix factorization (NeuCMF) framework for end-to-end recommendation.
• Apply contrastive learning principles during KG embedding learning to improve robustness and generalization, especially in the presence of noise in large-scale KGs.
• Use a two-stage training process: first pre-train KG embeddings via graph autoencoding, then jointly fine-tune with user-item interaction data under a mutual information regularization objective.

Experimental results

Research questions

• RQ1 Can knowledge graphs serve as an effective bridge to improve cross-domain recommendation by linking items across different domains through shared semantic knowledge?
• RQ2 How can item representations be jointly learned from user-item interactions and knowledge graphs to capture both domain-specific and domain-general knowledge?
• RQ3 What is the impact of incorporating higher-order neighborhood information from KGs on recommendation performance in sparse or cold-start scenarios?
• RQ4 How does mutual information maximization between user-item interaction embeddings and KG-based item embeddings improve cross-domain transfer learning?
• RQ5 To what extent does a KG-aware framework outperform existing CDR models that rely only on user overlap or embedding mapping functions?

Key findings

• The proposed KG-NeuCMF model achieves a 21% relative improvement in MAE over the strongest baseline on the movie-music cross-domain recommendation task, demonstrating significant performance gains.
• On the movie-book task, the model reduces MAE by 15.18% compared to the strongest baseline, highlighting its effectiveness in cross-domain settings with diverse domains.
• The model achieves the best F1_score of 90.80 ± 0.22 on the movie-book task, outperforming all baselines, including DDTCDR and CoNet, which are state-of-the-art CDR models.
• The ablation study shows that using a mutual information-based alignment between KG embeddings and user-item embeddings yields the best performance, outperforming simple concatenation or refinement via MLP.
• In cold-start scenarios, the model significantly outperforms NCF and CMF by leveraging KG-based representations for new items, reducing performance degradation due to data sparsity.
• The model’s performance is particularly strong in sparse domains like music, where KG integration alleviates data sparsity more effectively than in denser domains, confirming the value of KGs in low-data regimes.