PERSONALIZATION · 2026-10-26

Personalization — Paper Daily, 2026-10-26

Only one paper made the personalization track today: SynRec, on jointly training a recommender across businesses whose users do not overlap. Instead of shared user IDs, domains share some experts that learn domain-agnostic behavior rhythms, and each domain keeps its own experts. Teams with several business lines whose user data cannot be linked can start with its division of labor and its comparison across 6 domains.

3.50

1. SynRec: Synergistic multi-domain recommendation with frequency-guided expert specialization

When several businesses each run their own recommender and share no users, knowledge cannot be transferred between them through shared user IDs. This paper proposes SynRec, which lets several domains share one Transformer recommender and shares only domain-agnostic behavior patterns, such as the fast and slow rhythms of behavior sequences over time. It replaces the feed-forward layer with a domain-aware mixture of experts (MoE): some experts serve all domains and learn shared patterns, while others serve only one domain and learn its differences, which limits domains dragging each other down. Across 6 domains with no shared users, it beats the single-domain models in every domain and is the best of the evaluated multi-domain methods in 5 of them.