Abstract
Recommender Systems (RSs) are operated locally by different organizations in many realistic scenarios. If various organizations can fully share their data and perform computation in a centralized manner, they may significantly improve the accuracy of recommendations. However, collaborations among multiple organizations in enhancing the performance of recommendations are primarily limited due to the difficulty of sharing data and models. To address this challenge, we propose Decentralized Multi-Target Cross-Domain Recommendation (DMTCDR) with Multi-Target Assisted Learning (MTAL) and Assisted AutoEncoder (AAE). Our method can help multiple organizations collaboratively improve their recommendation performance in a decentralized manner without sharing sensitive assets. Consequently, it allows decentralized organizations to collaborate and form a community of shared interest. We conduct extensive experiments to demonstrate that the new method can significantly outperform locally trained RSs and mitigate the cold start problem.
Introduction
Recommender systems (RS) are the backbone of the modern web, yet they are frequently hampered by the "data sparsity" problem. Most users interact with only a negligible fraction of available items, making it difficult for local models to generate accurate predictions. While Cross-Domain Recommendation (CDR) addresses this by leveraging data from auxiliary domains, traditional approaches typically necessitate a centralized server where all data is pooled. In practice, data centralization is often unfeasible due to legal frameworks like GDPR/CCPA, ethical concerns, and commercial competition. Organizations require the benefits of a shared "brain" without compromising their private data.
Decentralized Collaboration Framework
To overcome these barriers, we propose a decentralized framework that allows organizations to form a "community of shared interest." Collaboration in the DMTCDR framework can occur in two primary configurations depending on the nature of the data overlap:
- User-Aligned: Organizations share a common user base but offer different sets of items (e.g., a movie streaming service and a book retailer).
- Item-Aligned: Organizations offer the same items but serve different user populations.
Methodology: MTAL and AAE
The core of our approach relies on two technical innovations designed to bridge the gap between disparate domains without centralizing data:
Multi-Target Assisted Learning (MTAL)
MTAL serves as a decentralized protocol for model improvement. Instead of sharing raw user profiles, each organization calculates "pseudo-residuals"—essentially a map of where their current local model is failing to make accurate predictions. These residuals are broadcast to peers, who use their private data and proprietary models to "fit" those residuals and return the resulting predictions. This iterative process allows all participating organizations to refine their local recommendations simultaneously.
Assisted AutoEncoder (AAE)
The AAE is a modified neural architecture designed for this decentralized setting. While standard autoencoders reconstruct local inputs, the AAE is trained to predict ratings or residuals across the global item set of all collaborating organizations, effectively leveraging insights from the entire community.
Experimental Evaluation
We evaluated the DMTCDR framework on three widely-used benchmark datasets: MovieLens1M (partitioned into 18 genre domains), Douban (Book, Movie, and Music domains), and Amazon (Books, Digital Music, Movies and TV, and Video Games).
Key Results
- Accuracy Boost: DMTCDR consistently outperformed models trained locally. We observed Root Mean Squared Error (RMSE) improvements of 3.8% on ML1M, 4.9% on Douban, and 6.7% on Amazon.
- Cold Start Mitigation: The framework proved highly effective for "cold start" scenarios, where a new organization with minimal data can immediately benefit from the collective intelligence of established peers.
- Privacy Preservation: The system maintains high recommendation accuracy even when integrated with Differential Privacy (DP) and Interval Privacy (IP) techniques, which add noise to the shared residuals to further protect sensitive data.
Conclusion and Future Work
DMTCDR demonstrates that decentralized collaboration can achieve performance gains comparable to centralized systems while maintaining strict data privacy. Future research will focus on optimizing hyperparameter sensitivity (specifically the gradient assistance weight), exploring more complex backbones such as Graph Neural Networks (GNNs), and ensuring equitable performance improvements across diverse domains through advanced loss function reweighting.
Summary updated