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.
The Challenge: Data Sparsity vs. Privacy
Modern recommendation engines face a fundamental "catch-22": high-quality suggestions require massive amounts of user interaction data, yet most individual platforms only observe a narrow slice of user behavior. While Cross-Domain Recommendation (CDR) attempts to bridge this gap by pulling data from related domains (e.g., using book ratings to predict movie preferences), traditional CDR is centralized. It assumes a single entity owns all data or that organizations are willing to pool sensitive user profiles and proprietary models—a scenario often impossible due to ethical, regulatory (like GDPR), and commercial constraints.
The DMTCDR Framework
To overcome these barriers, the authors introduce DMTCDR, a decentralized framework that enables organizations to boost recommendation accuracy without ever sharing raw data, models, or objective functions. The framework is designed to support two primary collaborative scenarios:
- User-Aligned: Different organizations share the same set of users (e.g., a bank and a movie streaming service).
- Item-Aligned: Different organizations rate or interact with the same set of items (e.g., different video platforms rating the same films).
Core Methodology
The system relies on two innovative components to facilitate decentralized collaboration:
Multi-Target Assisted Learning (MTAL)
MTAL serves as the decentralized "glue." Instead of sharing data, each organization calculates "pseudo-residuals"—mathematical maps indicating where their current local model is failing. These residuals are broadcast to peers, who use their own local data to "fit" the errors and send predictions back. Each organization then aggregates these peer insights to iteratively correct its own model. This allows every participant to improve simultaneously, unlike previous methods that focused on a single "sponsor" being helped by "assistors."
Assisted AutoEncoder (AAE)
The AAE is a modified neural network designed for decentralized tasks. It takes sparse input from a local domain but is architected to produce outputs that predict ratings across all participating domains. This allows the model to learn underlying relationships between different organizations' datasets without ever seeing the raw data itself.
Key Insights and Results
The researchers evaluated DMTCDR using major benchmarks including MovieLens 1M, Douban, and Amazon datasets. Key findings include:
- Superior Performance: DMTCDR consistently outperformed "Alone" baselines (training only on local data), with error rate (RMSE) improvements ranging from 3.8% to 12.7%.
- Competitive with Centralized Models: Surprisingly, the decentralized approach often performed as well as, or better than, "Joint" models where all data was pooled centrally. On the Douban dataset, DMTCDR achieved an RMSE of 0.857, significantly better than the Joint model's range of 0.899–1.081.
- Solving the Cold Start: DMTCDR is particularly effective for "cold start" organizations with very little data. By aligning with established peers, newcomers can "borrow" predictive power to serve their users far better than they could in isolation.
- Privacy Preservation: The system remains effective even when Differential Privacy (DP) is applied—adding noise to shared residuals to further mask sensitive information.
Conclusion and Future Work
DMTCDR demonstrates that organizations can form a "community of shared interest" to improve recommendation quality while maintaining strict data sovereignty. While the current framework is robust, future research will explore more diverse backbone models like Graph Neural Networks and methods to reduce the communication overhead of broadcasting residuals across rounds.
Summary updated