Optimization of Recall, Ranking, and Downward Modeling for the "Every Square Every House" Infinite-Scroll Light App

Series article summarizing a year of exploration and practice in recall, ranking, and cold‑start modules for the "Every Square Every House" light app; this is part six.

Scenario introduction: The light app is a content‑driven guide for home goods on Taobao, presenting scene‑based recommendations with multiple product anchors. Users can access it via keyword search, following the app, or clicking cards on the homepage. After entering the detail page, users can click product anchors to view product pages and perform actions such as collect or add to cart.

In early 2021 the detail page switched to an infinite‑scroll format, which provides smoother down‑scroll experience and increases exposure and consumption PV.

Recall optimization: To improve depth and click efficiency, multiple recall paths were added focusing on relevance, diversity, and hot fallback. Online 7‑day A/B showed significant lifts in uctr, pctr, down‑scroll depth, per‑user clicks and exposures.

Coarse ranking: Because the candidate set grew, a coarse‑ranking model was introduced to filter candidates before the heavy ranking stage, yielding modest gains in click metrics.

Sorting optimization – Multi‑objective learning: A multi‑task MMOE model was trained on first‑click, second‑click, and post‑click actions (collect, purchase) to boost multi‑step conversion.

Downward model: A lightweight model (W&D) was trained to predict whether a user will continue scrolling. Three versions of sample construction were evaluated (v1‑v3). Offline AUC and online A/B results showed v2 performed best overall.

Relevance modeling: Hero Content (the first clicked item) was used as context features and crossed with candidate items. Experiments v4‑v6 demonstrated up to 3.86% offline AUC improvement and consistent online gains, especially for the first 30 items.

Multi‑scene sample training: Combining samples from the main channel and a low‑traffic “first‑guess” scenario increased offline AUC by 4.63% and yielded modest online improvements.

Relevance strategy: Based on analysis of click‑through rates at different scroll positions, the system ensures that at least 30% of the first ten recommendations share attributes with the Hero Content, ranks items 11‑30 by preference scores, and prioritizes diversity and novelty thereafter.

Overall online impact: uctr +13.06%, pctr +6.01%, down‑scroll depth +4.70%, per‑user clicks +10.94%, exposures per view +3.96%.

Conclusion and outlook: The infinite‑scroll format brings higher exposure and consumption PV. Simple yet well‑targeted models and strategies can deliver strong gains; future work includes refining sample construction, enriching features, and further improving the down‑scroll model.

Acknowledgments: Thanks to mentors, algorithm and data teams, and the intelligent scenario team.

References: [1] https://dl.acm.org/doi/abs/10.1145/2988450.2988454

Team introduction: Alibaba Taobao Intelligent Team, a data‑and‑algorithm group serving multiple business lines, publishing research at KDD, ICCV, Management Science, etc.