STAN: A User‑Lifecycle‑Aware Multi‑Task Recommendation Model for Shopee

Introduction After the success of MMoE and PLE models, Shopee proposes a new multi‑task recommendation approach that incorporates user lifecycle information to further improve CTR, CVR, and stay‑time. The method is described in the paper (https://arxiv.org/abs/2306.12232).

Business Background Shopee’s feed consists of a two‑column live‑stream layout where users click a live video, spend time watching, and may purchase items. Users pass through distinct stages: new users with low orders and short stay, wandering users who browse longer but still have low conversion, and loyal users with higher CVR but shorter sessions.

Key Problems Existing multi‑task methods treat all users uniformly, causing a performance trade‑off (the “optimization seesaw”). The core challenge is to accurately capture user state so that CTR, stay‑time, and order metrics can be jointly optimized. This leads to three sub‑questions: how to identify user state, how to track it over time, and how to integrate it into a multi‑task model.

Solution – STAN STAN extends the traditional PLE/MMoE architecture with a left‑hand branch that models user information. An attention‑based feature extraction network generates user representations enriched with state‑related signals, which are incorporated into the loss function. A user‑adaptive Beta distribution is used to resample predictions when data are scarce, improving robustness.

Offline Evaluation – Understanding Experiments Visualization of user embeddings shows that STAN separates the three user groups (New, Wander, Loyal) more clearly than PLE. Temporal tracking demonstrates that the model can adaptively follow state transitions (e.g., from New to Wander).

Offline Evaluation – Shopee Dataset Using three weeks of training data and one week for testing, STAN achieves higher AUC and NDCG@1 than baseline PLE. Adding stage identifiers improves accuracy; incorporating adaptive stages and Beta resampling further stabilizes and lifts performance.

Offline Evaluation – Public Dataset Experiments on the WeChat Video Account dataset (tasks: like, click avatar, forward) confirm similar gains in AUC and NDCG@5, validating the method’s generality.

Online Value In production, STAN increases CTR by 3.94%, stay‑time by 3.05%, and orders by 0.88% compared with the PLE baseline, demonstrating practical impact despite the relatively small order volume on Shopee.

Conclusion and Outlook Modeling user lifecycle is essential for multi‑task recommendation. Future work will integrate lifecycle signals at finer granularity across all recommendation stages and continue to explore innovative methods that translate into business value.