Abstract
Recommendation systems often struggle with heterogeneous implicit feedback, such as distinguishing between browsing and purchasing behaviors. This paper introduces Role-based Bayesian Personalized Ranking (RBPR), a two-stage framework that models users as "browsers" to identify candidate items and then as "purchasers" to refine recommendations. By decoupling these roles, RBPR leverages plentiful but noisy browsing data alongside sparse, high-intent purchase data. Experimental results across five datasets, including large-scale ones like Netflix, demonstrate that RBPR achieves state-of-the-art accuracy while maintaining superior scalability compared to existing joint similarity learning methods.
Introduction
Most modern recommendation systems rely on "one-class" feedback—implicit signals like clicks or views where positive interactions are known, but negative ones are not. However, real-world user behavior is inherently heterogeneous. A user might browse dozens of items but only purchase a few. Standard models like Bayesian Personalized Ranking (BPR) often treat all interactions equally or ignore browsing data entirely, failing to capture the nuance between different levels of intent.
The challenge in Heterogeneous One-Class Collaborative Filtering (HOCCF) is twofold:
- Preference Nuance: Intent varies significantly between a "browse" and a "purchase."
- Scalability: Existing methods that attempt to integrate these signals often rely on complex joint similarity learning, which becomes computationally prohibitive for large-scale datasets like Netflix.
Methodology
RBPR addresses these challenges through a two-stage framework that mimics the natural shopping funnel. Instead of learning multiple behaviors simultaneously in a single complex model, it separates the user's intent into two sequential roles:
- Stage 1: Browser-based Preference Learning. The system treats the user as a "browser." A BPR model is trained on the union of all interactions (browses + purchases). This stage identifies items the user is likely to interact with at all, generating a candidate list of the top 3,000 items.
- Stage 2: Purchaser-based Preference Learning. The system then treats the user as a "purchaser." A separate BPR model is trained using only high-intent purchase data. This model re-ranks the candidates from Stage 1 to determine which items the user is most likely to actually buy.
Experimental Evaluation
The authors evaluated RBPR on five public datasets of varying scales: MovieLens 100K, 1M, 10M, Alibaba2015, and Netflix. For larger datasets where explicit "browse" vs. "purchase" labels were unavailable, the researchers simulated heterogeneous feedback by treating 5-star ratings as purchases and lower ratings as browses. This provided a rigorous test for how the model handles massive amounts of data.
Key Results
RBPR consistently outperformed standard BPR and other baselines across multiple metrics, including Precision, Recall, F1, and NDCG. For example, on the ML1M dataset, RBPR achieved an NDCG@5 of 0.1327 compared to BPR's 0.1121.
A critical finding was the model's scalability. Previous state-of-the-art methods like TJSL failed to complete training on the ML10M and Netflix datasets within 24 hours. In contrast, RBPR handled these large datasets efficiently and produced significantly better recommendations than the baseline.
Discussion and Conclusion
The success of RBPR lies in the "Sequential Role Assumption"—the realization that a purchaser is effectively a "converted" browser. By decoupling these roles, the model uses noisy browsing data to filter the search space and clean purchase data for high-precision ranking. This approach is not only more accurate but also significantly easier to implement using standard BPR building blocks.
Future work could involve expanding the number of roles (e.g., "searcher" or "rater") and testing on real-world datasets with native multi-action tracking, such as "add to cart" events, to further validate the approach.
Summary updated