Abstract
Platform-level federated recommendation systems face significant challenges due to "label skew," where different platforms exhibit varying rating distributions for the same items. This paper introduces FREIB (Federated Recommendation with Explicitly Encoding Item Bias), a framework designed to decouple intrinsic item characteristics from platform-specific rating behaviors. By explicitly encoding item bias and aligning features through global prototypes, FREIB significantly improves recommendation accuracy and consistency across decentralized platforms.
Introduction: The Challenge of Label Skew
In platform-level federated recommendation, multiple organizations (e.g., streaming services) collaborate to train a model without sharing raw user data. A primary hurdle is label skew, which manifests in two ways:
- Rating Bias: Platforms have different "rating habits." Users on one platform might be consistently more critical than those on another.
- Item Bias: Items possess intrinsic qualities (e.g., a high-budget masterpiece vs. a low-quality production) that should be recognized universally.
Existing models often conflate these biases, leading to "fuzzy" item embeddings and poor generalization. When a strict platform provides a low rating and a lenient platform provides a high rating for the same item, conflicting gradients hinder the learning process.
The FREIB Framework
To address these issues, the authors propose the FREIB framework, which augments the standard Neural Collaborative Filtering (NCF) backbone with three innovative components:
1. Explicit Item Bias Encoder (EIBE)
FREIB introduces a dedicated encoder to capture the "intrinsic worth" of an item independently of user-platform interactions. This separation allows the model to learn stable item features that are not distorted by local rating distributions.
2. Global Knowledge Guidance (GKG)
To maintain alignment with a global consensus, the server-side model acts as a teacher. A distillation loss ($L_{distill}$) is employed during local training, encouraging local models to adhere to global rating patterns and preventing them from drifting too far into platform-specific skews.
3. Feature Prototype Alignment (FPA)
FPA regularizes local training by ensuring that items with similar ratings share similar bias features across the federation. Local "prototypes" (mean feature vectors for each rating level) are aggregated into global prototypes at the server. These global anchors help platforms with sparse data learn what high-quality item embeddings should look like.
Experimental Evaluation
The researchers evaluated FREIB on three public datasets: MovieLens-100K, MovieLens-1M, and Amazon-Beauty. To simulate real-world platform-level federation, they used Dirichlet sampling to distribute data across five participants with varying degrees of label skew.
Key Results
- Superior Accuracy: FREIB consistently outperformed state-of-the-art baselines like FedProx and PFedRec. On the highly skewed MovieLens-100K dataset, FREIB achieved an MAE of 0.7369, significantly better than the next best baseline (0.9371).
- Stability and Convergence: The explicit modeling of item bias leads to faster and steadier convergence compared to methods that struggle with conflicting gradients.
- Privacy Preservation: The framework remains robust even when integrated with Local Differential Privacy (LDP), ensuring data protection without sacrificing significant performance.
Conclusion and Future Work
FREIB demonstrates that decoupling item bias from rating bias is essential for effective platform-level federated recommendation. While the framework introduces some computational overhead due to the additional encoders and prototype alignment, the gains in accuracy and stability are substantial. Future research will focus on optimizing these overheads and exploring the semantic interpretability of the learned bias prototypes to better understand the underlying factors of item quality.
Summary updated