FedCGR: Federated Cross-Domain Generative Recommend...

FedCGR: Federated Cross-Domain Generative Recommend... | AI Research

Key Takeaways

FedCGR is a federated learning framework designed to improve cross-domain recommendation (CDR) by treating it as a generative task.
To address this tension, we revisit federated CDR as generation over a stable semantic item language.
To overcome these constraints, we propose FedCGR, a federated generative CDR framework that keeps the item language stable and makes adaptation explicit.
It addresses the challenge of aligning item spaces across different platforms without sharing private user interaction data.
By using a shared, stable vocabulary of item descriptions, the framework allows multiple domains to collaborate on training a recommendation model while keeping sensitive behavioral data local to each client.

Paper AbstractExpand

Cross-domain recommendation (CDR) transfers preference knowledge across related domains, but federated deployment makes cross-domain alignment difficult because the behavioral anchors that align item spaces, such as overlapping users and shared interaction signals, are often sparse, unavailable, or privacy-sensitive across clients. To address this tension, we revisit federated CDR as generation over a stable semantic item language. By representing items as discrete semantic ID (SID) sequences derived from public item-side metadata, cross-domain item alignment is induced by a shared vocabulary rather than by exchanging private interactions or aligning domain-specific embeddings. Directly federating SID-based generators, however, introduces two design constraints: the SID tokenizer must remain fixed to preserve cross-client token consistency, which creates a semantic-only bottleneck because local collaborative filtering (CF) signals cannot be globally shared or aligned; meanwhile, standard federated averaging can cause negative transfer under domain heterogeneity. To overcome these constraints, we propose FedCGR, a federated generative CDR framework that keeps the item language stable and makes adaptation explicit. FedCGR injects local CF evidence through a reliability-aware semantic interface and trains a prototype-personalized generator that selectively aggregates shared parameters according to domain relatedness while keeping domain-specific quantities local. Experiments on six Amazon cross-domain scenarios show that FedCGR consistently outperforms federated generative baselines and achieves competitive performance against strong sequential and federated CDR methods under both full-ranking and sampled evaluation protocols.

FedCGR is a federated learning framework designed to improve cross-domain recommendation (CDR) by treating it as a generative task. It addresses the challenge of aligning item spaces across different platforms without sharing private user interaction data. By using a shared, stable vocabulary of item descriptions, the framework allows multiple domains to collaborate on training a recommendation model while keeping sensitive behavioral data local to each client.

Solving the Federated Alignment Problem

Traditional CDR methods often rely on sharing user interactions or aligning continuous item embeddings, which is difficult in federated settings where data is siloed and privacy is a priority. FedCGR avoids this by representing items as discrete Semantic ID (SID) sequences derived from public metadata. Because these SIDs are generated from a shared vocabulary, they provide a common language for items across different domains, removing the need to exchange private interaction signals to achieve alignment.

The FedCGR Architecture

The framework relies on two primary mechanisms to ensure effective knowledge transfer:

Reliability-Aware Semantic Interface: Each client uses a fixed SID tokenizer to represent items. To account for domain-specific preferences, clients extract local collaborative filtering (CF) evidence from their private data. This evidence is injected into the model as a residual signal, weighted by an item-level reliability score that prioritizes high-confidence interactions and suppresses noisy, long-tail data.
Prototype-Personalized Generator: The recommendation model uses a shared-private Mixture-of-Experts (MoE) architecture. While some model parameters are shared across all domains, others remain local. The server aggregates shared parameters based on "domain prototypes"—compact summaries of domain behavior—ensuring that knowledge is transferred more strongly between domains that are behaviorally similar.

Performance and Evaluation

The authors evaluated FedCGR across six Amazon cross-domain scenarios. The results indicate that the framework consistently outperforms existing federated generative baselines. Furthermore, it achieves competitive performance against established sequential and federated CDR methods under both full-ranking and sampled evaluation protocols. The researchers note that the effectiveness of the model depends on the balance between CF enrichment and personalized aggregation, which varies depending on how closely related the domains are.

Key Considerations

While FedCGR provides a way to align item spaces, it operates under specific design constraints. The SID tokenizer must remain fixed after initial training to maintain consistency across clients, which creates a "semantic-only" bottleneck that the framework attempts to mitigate through local CF injection. Additionally, the authors clarify that while the framework protects raw interaction sequences, it does not inherently provide formal differential privacy; however, they note that standard privacy-preserving techniques like secure aggregation could be integrated with the system.