FedMM: Federated Quantization Boosts Multi-Market CTR Prediction While Preserving Privacy

Online platforms like Amazon and Netflix serve users across multiple countries, making multi-market recommendation (MMR) critical. Traditional MMR methods use a pre-train-and-fine-tune paradigm on centralized data, which ignores privacy. Federated learning offers privacy but fails with disjoint ID spaces and market heterogeneity. To address this, researchers from multiple institutions propose FedMM (Federated Collaborative Signal Quantization). The core idea is to use a residual quantized variational autoencoder (RQ-VAE) with a hierarchical dual-layer codebook. The first layer is a global federated codebook updated via aggregation to capture cross-market shared collaborative patterns. The second layer is a local codebook that learns market-specific semantics. This quantization approach preserves privacy by transmitting discrete codes instead of raw embeddings and aligns disjoint ID spaces across markets.

FedMM is evaluated on benchmark datasets for CTR prediction, showing significant improvements over existing methods while maintaining strict privacy guarantees. The method integrates learned discrete codes—combining general and specific collaborative signals—into downstream CTR models to enhance accuracy across all markets. The work has been accepted at SIGIR 2026, a top venue for information retrieval research. FedMM's approach offers a practical path for large-scale recommendation systems that must balance performance with data privacy regulations across different regions, and its codebook mechanism could inspire further research into federated learning for heterogeneous domains.