Google's HA-MoE improves Discover feed ranking with heterogeneous AI

Google Discover's feed blends web articles, short and long-form videos, and user-generated content from the open web, creating a ranking challenge far beyond homogeneous music or video platforms. Each content type has different feature densities and interaction patterns, and unified ranking models often suffer from majority bias or negative transfer. In a new ACM RecSys 2026 case study, Google researchers propose HA-MoE, a multi-gated mixture-of-experts architecture that injects explicit heterogeneity context into both the gating networks and expert representations. This lets the model specialize effectively without adding significant operational overhead, a critical factor at industrial scale.

To make the system reliable, they also introduce LENS, a lightweight observability framework that diagnoses expert specialization and tracks how that functional heterogeneity evolves during continuous retraining. Evaluation uses DL-AUC, a metric that combines global ranking performance with cross-segment correctness. Offline experiments on large-scale industrial datasets show consistent gains over baselines, and online A/B tests confirm real-world improvements in feed activity and exploration. The work validates a practical path for handling heterogeneous feeds in production settings.