Scenario-Adaptive and Self-Supervised Multi-Scenario Personalized Recommendation (SASS)

The paper introduces the problem of multi‑scenario modeling in personalized recommendation, where each recommendation entry (scene) shares the same feature and label space but exhibits different data distributions, leading to data sparsity and high maintenance costs.

Four key challenges are identified: precise scene‑wise information transfer, leveraging unsupervised data to alleviate sparsity, applying multi‑scenario modeling to the recall stage, and reducing iteration and deployment costs.

The proposed solution, named SASS, consists of two stages. Stage 1 performs self‑supervised pre‑training using contrastive learning on unlabeled data, treating a user's interactions in different scenes as natural augmentations to align scene representations. Stage 2 fine‑tunes the model on labeled click (or watch‑time) data for the recall task, reusing the pre‑trained embeddings and network.

The model architecture features a shared embedding layer, a global shared network, scene‑specific subnetworks, and a bias network. Adaptive gate and update gate units control how much global information is transferred to each scene, enabling fine‑grained information migration.

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

Ablation studies validate the effectiveness of the adaptive gating mechanism, the benefit of the pre‑training task, the contribution of the bias network, and the impact of network depth.

Online A/B tests in Taobao’s content recommendation pipelines (short video and image‑text) demonstrate significant gains in hit‑rate and NDCG, confirming the model’s practical impact.

The article concludes with a Q&A covering evaluation metrics, feature alignment across scenes, code open‑source plans, negative sampling, deployment details, and future directions such as cross‑domain extensions.