Deep Learning Applications in Alibaba 1688 B2B E‑commerce Recommendation System: From Deep Match to Live Content Ranking

The presentation by senior algorithm expert Karbon Ali from Alibaba explores the application of deep learning in the recommendation system of the B2B e‑commerce platform 1688, covering product recall (Deep Match), ranking (DIN, DMR) and live‑content recommendation.

It outlines the system's evolution from 2017 to 2019: starting with heuristic I2I/E‑TREC methods, moving to WDL, adopting YouTube‑inspired Deep Match for multi‑interest recall, and later integrating DIN, DIEN, and multi‑tower models for ranking.

Product recall is described through three approaches—I2I recall, U2I recall, and Deep Match U2I recall—detailing their mechanisms, advantages such as real‑time triggering and high coverage, and challenges like interpretability.

Practical lessons for Deep Match include using random negative sampling, effective position embeddings, and enriching item feature sequences, which together yielded notable online gains in click‑through and exposure metrics.

Ranking models are examined in depth: DIN introduces attention‑based user representation, achieving CTR +5% and CVR +11% over baseline WDL; LSRMM adds long‑term interest modeling and raw feature embeddings, improving CTR by 1.57%; DMR further incorporates user‑to‑item relevance, delivering additional lifts in UV‑CTR, per‑user clicks, and conversion.

The live‑content recommendation segment discusses business motivations, multi‑objective learning challenges, and the transition from a feature‑engineered LR/GBDT model (V1) to a deep double‑sequence multi‑task model (V2) and finally to a heterogeneous‑network (HIN) attention model (V3), which leverages shared and dedicated embeddings and pre‑trained metapath2vec vectors, achieving significant gains in CVR, stay time, and overall engagement.

Future directions include deploying twin‑tower coarse‑ranking, multi‑task learning with MMOE, and further HIN pre‑training to enhance both recall and ranking, as well as exploring emerging content‑understanding techniques.

The article concludes with contact information for the 1688 recommendation team and invites readers to join the DataFunTalk community.