Adaptive Domain Interest Network for Multi-domain Recommendation

Industrial recommendation systems often serve multiple business scenarios, requiring scenario-specific adjustments during the recall stage. Traditional approaches either train separate models per scenario, ignoring cross‑scenario user and item relationships, or train a single shared model, which struggles to capture significant inter‑scenario differences.

We propose the Adaptive Domain Interest Network (ADIN), which adaptively handles commonalities and divergences across scenarios. ADIN consists of a shared backbone and scenario‑private subnetworks, a domain‑specific batch normalization (DSBN) layer, and a domain‑interest adaptive layer based on SE‑Block attention. A self‑supervised training strategy further captures label‑level correlations between scenarios.

Extensive experiments on both public and industrial datasets demonstrate that ADIN consistently outperforms baseline models. Ablation studies confirm the effectiveness of DSBN, the domain‑interest adaptive layer, and the self‑supervised component. Online A/B tests in Alibaba’s display‑ad platform show a 1.8% increase in revenue and improvements in RPM, PPC, and CTR across three scenarios.

The work validates that domain adaptation techniques can significantly enhance multi‑scenario recommendation at the recall stage, and ADIN has been fully deployed in production.