Daily Good Shop: Two‑Stage Card Ranking and Multi‑Task Modeling for E‑commerce Recommendations

Daily Good Shop is a unique shop‑guidance scenario aiming to help users discover more high‑quality stores. The main entry points include the home page grid and the information flow, where users click on selected shop cards and are directed to secondary theme pages.

The card layout consists of a two‑level structure: a shop (store) and multiple products. Card ranking is modeled as a cascade of two sub‑models: product ranking (to select items for a card) and card ranking (to order the assembled cards).

Two‑Stage Modeling : The ItemAwareShopRankingModel (IASM) first predicts product click‑through rates (CTR) using long‑term user behavior models such as SIM and ETA. The top‑3 products are then assembled into a card, and a card‑level model predicts the overall CTR.

Product Ranking : Long‑term user behavior sequences are incorporated via SIM and ETA architectures. Due to limited data, only the LogitsLayer and BiasNet of the pretrained model are fine‑tuned.

Multi‑Task Modeling : Three objectives are jointly optimized—click, subscription, and lead‑IPV. Click and subscription are modeled as binary classification, while lead‑IPV is treated as a multi‑class problem after equal‑frequency bucketing. A shared‑bottom network shares low‑level parameters, with separate towers for each task.

Two fusion strategies are explored:

Formula‑based fusion extending CTR/CVR combination to click, subscription, and lead‑IPV scores.

Learning‑to‑Rank (LTR) fusion using a GBDT model to fit a weighted combination of task scores, yielding notable gains in exposure and click metrics.

Card Re‑ranking : Beam‑search is used to generate candidate card sequences, selecting the sequence with the highest page‑level IPV value. The model also predicts whether a user will scroll down (down‑flip) as an auxiliary task.

Online A/B tests show significant improvements: +2.65% in average clicks per user, +34.29% in subscription, +32.88% in lead‑IPV, and +21.56% in dwell time after deploying the shared‑bottom multi‑task model. Card re‑ranking further increases exposure and click metrics.

Future Work includes exploring global optimal card assembly (e.g., one‑shop‑multiple‑cards), list‑wise re‑ranking, and joint ordering of products within cards.