Abstract
Traditional recommendation systems often fail to capture the multifaceted nature of user interests by representing each user as a single point in a latent space. This paper introduces Diversity-Promoting Collaborative Metric Learning (DPCML), a framework that assigns multiple embedding vectors to each user. By integrating an Adaptive Preference Assignment (APA) strategy with a novel differentiable negative sampling technique (DiHarS), DPCML effectively captures diverse user preferences, achieving state-of-the-art performance in both recommendation accuracy and diversity across multiple real-world datasets.
The Challenge of Unimodal User Representations
Most collaborative filtering models rely on the "one user, one vector" paradigm. While computationally efficient, this approach assumes that a user's interests are unimodal. In practice, a user might have distinct, unrelated interests—such as a preference for both high-octane Sci-Fi and niche Horror films.
Standard Collaborative Metric Learning (CML) attempts to pull a single user vector toward all items they have interacted with. When a user's interests are imbalanced (e.g., 90% Sci-Fi and 10% Horror), the model naturally biases toward the majority interest, often failing to recommend items from the minority category. This "preference bias" limits the system's ability to provide truly diverse and personalized recommendations.
The DPCML Framework
DPCML addresses these limitations through four primary technical innovations:
- Multi-vector User Representation: Each user is represented by a set of vectors. The preference score for an item is determined by the minimum distance between that item and any of the user's vectors, allowing different vectors to specialize in different interest clusters.
- Adaptive Preference Assignment (APA): Recognizing that not all users are equally diverse, APA dynamically determines the number of vectors for a user based on their interaction history. Users with more interactions are statistically more likely to possess varied interests.
- Diversity Control Regularization Scheme (DCRS): To ensure the multi-vector representation remains effective, DCRS prevents vectors from collapsing into a single point or drifting too far apart, maintaining a "safe" range of internal diversity.
- Differentiable Hardness-aware Sampling (DiHarS): The authors introduce DiHarS to optimize for the top-$U$ hardest negative examples. They demonstrate that standard hard negative sampling is limited, and a more nuanced, differentiable approach to "hardness" significantly improves ranking performance.
Empirical Evidence and Results
The model was rigorously tested on six datasets, including MovieLens, Steam, and CiteULike. Analysis of the MovieLens datasets reveals a broad distribution of user diversity, justifying the need for multi-vector representations.
Key findings from the experiments include:
- Superior Accuracy: DPCML consistently outperformed established baselines like NeuMF and LightGCN. For instance, on the Steam-200k dataset, DPCML achieved an MRR of 54.90, compared to 40.10 for standard CML.
- Enhanced Diversity: Beyond accuracy, the model showed significant gains in Intra-List Similarity (ILS) and Coverage, indicating a more varied and representative set of recommendations.
- Generalization: Theoretical analysis using covering numbers suggests that DPCML can achieve lower generalization error than traditional CML, despite the increased parameter count.
Conclusion
DPCML represents a significant step forward in making recommendation systems more representative of real-world user behavior. By moving to a multi-vector paradigm and refining the sampling process, the framework successfully balances the competing demands of accuracy and diversity. Future work may explore end-to-end learning of the optimal number of interest clusters and deeper integration of semantic item metadata.
Summary updated