ATLAS: Learning to Recommend Across Unseen Domains

Modern recommender systems are typically domain-bound, meaning they require retraining or fine-tuning when applied to new product categories or catalogues. The authors investigate whether a single, frozen recommendation model can be trained on multiple heterogeneous source domains to generalize effectively to entirely unseen domains—a setting they term Recommendation Domain Generalization (RDG). Unlike cross-domain recommendation, which assumes the target domain is available during training, RDG requires the model to perform zero-shot inference without any parameter updates.

To achieve this, the authors introduce ATLAS, a framework that constructs a unified user-item retrieval space using three key mechanisms:

1. Semantic and Adversarial Alignment: Item representations are mapped into a shared space using a frozen text encoder and a domain-adversarial objective, which forces the model to discard domain-specific features.
2. Geometric Alignment: User populations are aligned using the Gromov-Wasserstein lower bound, which preserves the intrinsic structure of user-item interactions across domains without requiring overlapping users or items.
3. Residual Vector Quantization (RVQ): User and item embeddings are compressed into a discrete latent space using hierarchical codebooks. This acts as an information bottleneck, suppressing residual domain-specific noise and ensuring the model focuses on transferable interaction patterns.

ATLAS was trained on five Amazon domains and tested on ten unseen domains. It consistently outperformed state-of-the-art baselines, including sequential, graph-based, and LLM-based recommenders, achieving an average relative gain of 24% in HitRate. The authors also identified a "source-domain diversity effect," where increasing the heterogeneity of the training data significantly improved the model's ability to generalize to new, unseen environments.