Scenario-Adaptive and Self-Supervised Multi-Scenario Personalized Recommendation (SASS): Design, Training, and Deployment

The paper introduces the problem of multi‑scenario modeling in recommendation systems, where each "scenario" corresponds to a different entry point or page (e.g., homepage, product detail, short‑video feed). Data distributions differ across scenarios, leading to sparsity and high maintenance costs.

Four key challenges are identified: (1) fine‑grained information transfer between scenarios, (2) leveraging unsupervised data to alleviate sparsity, (3) reducing iteration and operation costs, and (4) applying multi‑scenario modeling to the recall stage of the recommendation pipeline.

Four baseline solution families are discussed: (a) independent models per scenario, (b) a single model trained on mixed data, (c) a two‑stage pre‑train + fine‑tune approach, and (d) joint training with shared and scenario‑specific subnetworks. The authors adopt the fourth family.

The proposed SASS framework consists of two stages. Stage 1 performs self‑supervised pre‑training using contrastive learning (InfoNCE loss) on unlabeled user‑item interactions across paired scenarios, encouraging scenario‑aware embeddings. Stage 2 fine‑tunes the model on labeled click data for the recall task, reusing the pre‑trained embeddings and network structure.

The model architecture includes a shared embedding layer, a global shared network, and per‑scenario expert networks, plus a bias network that injects scenario‑specific features. Adaptive and update gates control how much global information is transferred to each scenario.

Extensive experiments on two public datasets and Alibaba’s industrial data compare SASS‑Base (without pre‑training) and SASS (with pre‑training) against single‑scenario models, mixed‑sample models, and other multi‑scenario methods. Results show that joint modeling consistently outperforms baselines, and the self‑supervised pre‑training further improves performance, especially in sparse or small scenarios.

Ablation studies validate the effectiveness of the adaptive gate, the contribution of the pre‑training task, the auxiliary global loss, and the depth of the transfer network. Online A/B tests in Taobao’s content recommendation (short video and image‑text) confirm significant metric gains, particularly for low‑traffic scenarios.

In summary, the SASS approach provides a unified framework that mitigates data sparsity, reduces model maintenance overhead, and is successfully deployed in production for multi‑scenario recall, demonstrating the practical value of scenario‑adaptive and self‑supervised learning in large‑scale recommendation systems.