Abstract
Cross-domain recommendation (CDR) leverages data from multiple platforms to improve suggestion accuracy. Traditional models often assume shared features have identical latent representations across domains, ignoring platform-specific nuances. We present Coupled Factorization Machines (CoFM), which decompose latent factors into shared and domain-specific components. Experiments on MovieLens and Netflix datasets demonstrate that CoFM significantly outperforms state-of-the-art baselines by capturing both commonalities and unique domain characteristics.
Introduction
Cross-domain recommendation (CDR) aims to improve suggestions in one domain (e.g., Netflix) by leveraging data from another (e.g., MovieLens). Most existing CDR models assume that if two datasets share a field—like "Movies"—the latent factors representing those movies should be identical across both platforms.
However, this assumption is often flawed. For instance, MovieLens allows ratings in 0.5 increments, while Netflix uses whole integers. Furthermore, user behavior and platform demographics differ between services. Forcing these different "data dialects" into a single shared latent space ignores domain-specific uniqueness, which can lead to sub-optimal recommendations. The challenge is to exploit the correlation between domains without losing the nuances that make them different.
Methodology: Coupled Factorization Machines
The researchers proposed Coupled Factorization Machines (CoFM), which modify the latent vectors of coupled fields. Instead of a single latent vector for a shared feature, CoFM splits the vector into two parts:
- The Shared Part ($c$): Captures characteristics common to both domains (e.g., the inherent "action-ness" of a movie).
- The Domain-Specific Part: Preserves the uniqueness of each domain (e.g., how "action-ness" is perceived or rated specifically on Netflix vs. MovieLens).
The objective function calculates interactions by combining these shared and non-shared elements. This allows the model to learn a "global" understanding of items while maintaining a "local" calibration for each specific platform. A hyperparameter $c$ allows practitioners to tune the balance between shared and domain-specific information.
Experimental Results
The model was evaluated using MovieLens and Netflix datasets, focusing on approximately 5,911 movies appearing in both. CoFM was compared against several baselines, including standard Factorization Machines (FM), Collective Matrix Factorization (CMF), and deep learning-based models like DSSM.
As shown in the results table, CoFM consistently achieved the lowest Root Mean Square Error (RMSE) across both datasets. These performance gains held true across various dimensions of latent spaces ($k$ values ranging from 5 to 25). Statistical significance tests (t-tests) confirmed that CoFM’s improvements over existing state-of-the-art models were statistically significant, with p-values well below $0.05$.
Conclusion
CoFM provides a mathematical middle ground for coupled data that is neither entirely identical nor entirely independent. By explicitly accounting for domain-specific variations in rating scales and user behavior, it achieves superior recommendation accuracy. Future research directions include extending CoFM to higher-order feature interactions and utilizing deep neural networks to learn more complex mappings between domains.
Summary updated