Improving New User Experience in Taobao Live Recommendation via Multi‑Channel Lifelong Product Sequence Modeling

The article analyzes the cold‑start problem of new users in Taobao’s live‑streaming e‑commerce platform. While high‑activity users generate abundant interaction data, new users exhibit sparse, random visits, leading to low exposure and conversion rates.

Key challenges include real‑time recommendation constraints (only currently live streams can be shown), the multimodal nature of live content (cover images, titles, ASR transcripts, host attributes, product information), and the difficulty of aligning cross‑domain signals such as user purchase history with live‑stream features.

To address these issues, the authors propose a multi‑channel lifelong product‑sequence network. User purchase sequences are enriched with side‑information (weights, position, stay time, shop/brand/category IDs) and processed through five extraction channels: three relevance‑based channels driven by the target live‑stream, one user‑interest channel, and one auxiliary channel. Each channel selects the most relevant subsequence via top‑K scoring, and the results are merged.

The overall model architecture combines classic recommendation columns (user, item, context, host) with the extracted sequences using DIN‑style target attention. Two towers are employed: a Main Net for primary CTR scoring and a Bias Net to capture user‑specific bias toward certain hosts. An auxiliary loss further calibrates user interest representations derived from lifelong sequences.

Training optimizations include initializing the new sequence attention weights with pretrained target‑attention parameters (Mixout regularization) and applying channel‑wise softmax to mitigate probability dilution caused by long merged sequences.

Offline experiments show that the multi‑channel approach improves rtpScore from 0.647 (baseline) to 0.72–0.825, confirming stronger relevance extraction. Online A/B tests report significant lifts in first‑page UV/PV, second‑page click‑through, dwell time, and next‑day return rates, especially for low‑activity and re‑engaged users.

Future work plans to extend the framework to other live‑stream scenarios (mixed product‑card feeds, vertical scrolling), add diversity/novelty channels, and explore adversarial generation of user representations.