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

Introduction Since the success of MMoE and PLE models, Shopee proposes a new multi‑task recommendation approach—STAN—that refines user segmentation based on lifecycle stages to improve CTR, CVR, and stay‑time.

Business Background Shopee’s feed consists of a dual‑column layout where users progress through stages: new users with low orders and time, wandering users with increasing dwell time but low conversion, and loyal users with high CVR but shorter sessions. Different user groups exhibit distinct preferences for recommendation tasks.

Key Problems Existing multi‑task methods treat all users uniformly, causing a trade‑off (optimization seesaw). The core issues are: (1) how to identify user state, (2) how to accurately track user state over time, and (3) how to incorporate state information into multi‑task models.

Solution – STAN STAN augments the traditional PLE/MMoE architecture with a left‑hand side that models user information. An attention‑based user feature extractor generates state‑aware embeddings, which are used to adjust the loss. Three sub‑components address the key problems:

Identify user state via attention‑driven user representations derived from interaction features.

Track user state by estimating per‑task scores and applying a user‑adaptive Beta‑distribution resampling to mitigate uncertainty for sparse data.

Combine state with multi‑task learning by adding a state‑aware loss term to the existing multi‑task loss, training jointly.

Offline Evaluation Understanding experiments show STAN’s embeddings separate Wander, Stick, and Loyal users better than PLE. On Shopee’s internal dataset, STAN improves AUC and NDCG@1 across tasks, especially after adding adaptive stages and Beta resampling. Similar gains are observed on a public WeChat video dataset (AUC, NDCG@5).

Online Results Deploying STAN in production (baseline PLE) yields CTR + 3.94 %, stay‑time + 3.05 %, and order + 0.88 % improvements, confirming the model’s effectiveness in real‑world traffic.

Conclusion & Outlook Modeling user lifecycle is crucial for multi‑task recommendation. Future work will integrate finer‑grained lifecycle adjustments at each candidate generation stage and continue to innovate methods that translate into business impact.