Abstract
Cross-Domain Recommendation (CDR) is a powerful technique for addressing the cold-start problem by transferring knowledge from a data-rich source domain to a sparse target domain. However, the assumption that all source domain interactions are beneficial often leads to the inclusion of "noise"—irrelevant user-item relationships that can degrade recommendation performance. This paper introduces a three-layer framework designed to refine graph representations through targeted edge pruning. By projecting domain-specific embeddings into a shared latent space and applying anomaly detection, the proposed method identifies and removes outlier edges. Evaluations on the Amazon dataset demonstrate that this selective "forgetting" of irrelevant data leads to significant improvements in recommendation accuracy, outperforming existing state-of-the-art models.
Introduction
Recommendation systems frequently encounter the cold-start problem, where a lack of historical data for new users or items hinders accurate suggestions. Cross-Domain Recommendation (CDR) mitigates this by utilizing data from a related source domain (e.g., using book purchase history to recommend movies). A significant challenge in CDR is the presence of irrelevant source domain data. For instance, a user's specific interest in technical manuals may not correlate with their movie preferences. Existing CDR methods often fail to distinguish between useful transfer knowledge and noise. This research proposes a method to identify and eliminate these unnecessary node interactions to improve target domain recommendations.
Methodology
The proposed framework consists of three distinct layers: the Embedding Calculation Layer, the Edge Pruning Layer, and the Prediction Layer.
1. Embedding Calculation Layer
Using LightGCN, a state-of-the-art Graph Convolutional Network, the system learns separate embedding features for users and items in both the source and target domains based on purchase histories. This provides the initial high-dimensional representation of the user-item graph.
2. Edge Pruning Module
This core innovation addresses the noise problem through a two-step process in a shared latent space.
- Latent Space Projection: Since embeddings from different domains are not directly comparable, Canonical Correlation Analysis (CCA) is employed to project them into a shared space where they can be aligned.
- Anomaly Detection and Pruning: The Local Outlier Factor (LOF) algorithm identifies source-domain items that are "outliers" relative to items in the target domain. Edges connected to these outliers are pruned, effectively cleaning the graph of irrelevant information before the final recommendation step.
3. Prediction Layer
The refined, pruned graph is processed by a Preference Propagation GraphNet (PPGN) to generate the final item recommendations for the target domain.
Experimental Results
The researchers evaluated their method using the Amazon review dataset, specifically focusing on the Books domain as the source and Movies/TV as the target. The dataset included 37,388 mutual users who had interacted with both domains.
Key findings include:
- Superior Accuracy: The pruned model consistently outperformed baselines across several metrics. At $k=5$, the proposed method achieved an NDCG of 0.428, compared to 0.421 for the same model without pruning and 0.015 for standard LightGCN.
- Optimal Pruning Ratio: The best results were achieved when the number of neighboring items ($v$) was set to 5, which corresponded to pruning roughly 3.55% of the edges in the source domain. Pruning too much or too little resulted in lower performance, suggesting a specific "sweet spot" for noise removal.
- Capturing Nuance: Qualitative analysis showed that while baseline models often recommend items based on broad genres, the pruning method was better at capturing nuanced preferences, such as specific character archetypes or narrative styles.
Conclusion
This research demonstrates that "less is more" in the context of Cross-Domain Recommendation. By intentionally pruning noisy edges in a shared latent space, the proposed framework achieves superior accuracy on sparse datasets. While the method introduces some computational complexity, the gains in recommendation quality are substantial. Future work will focus on optimizing the hyperparameter tuning process and exploring the method's performance on more disparate domain pairs.
Summary updated