LiveForesighter: Generating Future Information for Live-Streaming Recommendations at Kuaishou

TL;DR

LiveForesighter tackles the challenge of live-streaming recommendations by predicting future information to guide both moment-level distribution and near-future engagement. The method combines statistic-sequence modeling to detect high-light moments with product-sequence forecasting to anticipate future user interest, all within a Transformer-based generative framework that is disentangled from the core ranking model. Empirical results from two large-scale Kuaishou services show consistent offline AUC gains and meaningful online improvements in GMV, order times, and viewing metrics, validating the practical impact of forecasting future signals. The work demonstrates a practical pathway for incorporating future information into live-streaming RecSys and points to future enhancements via multi-modal, pre-trained language models.

Abstract

Live-streaming, as a new-generation media to connect users and authors, has attracted a lot of attention and experienced rapid growth in recent years. Compared with the content-static short-video recommendation, the live-streaming recommendation faces more challenges in giving our users a satisfactory experience: (1) Live-streaming content is dynamically ever-changing along time. (2) valuable behaviors (e.g., send digital-gift, buy products) always require users to watch for a long-time (>10 min). Combining the two attributes, here raising a challenging question for live-streaming recommendation: How to discover the live-streamings that the content user is interested in at the current moment, and further a period in the future?

Figures (5)

• Figure 1: Live-streaming high-lights could be reflected by the crowd of users behaviours.
• Figure 2: The interpreting product sequence of a online-shopping live-streaming.
• Figure 3: An overview of model architecture of LiveForesighter.
• Figure 4: LiveForesighter Deployment.
• Figure 5: Live-streaming exposure trend across time.