Evolution of Large‑Scale Recommendation Models at Weibo: Technical Roadmap and Recent Advances

The presentation outlines Weibo's recommendation platform, describing the diverse business scenarios (home‑page tabs, hot‑search streams, immersive video) and the challenges of high‑traffic, multi‑modal content with varied user feedback.

A technical roadmap is shown, highlighting the transition from early FM‑based models to deep, real‑time architectures, the in‑house Weidl online‑learning platform, and the ability to switch back‑ends quickly.

Recent large‑model iterations focus on multi‑objective fusion (static, RL‑based, and model‑driven weighting), multi‑task learning (MMOE → SNR → DMT), and multi‑scenario techniques (slot‑gate layers) to handle heterogeneous goals such as click, dwell time, interaction, and completion.

Interest representation advances from DIN to SIM/DMT, with longer behavior sequences and ultra‑long user histories improving personalization, especially for low‑exposure items.

Feature engineering discusses the impact of massive ID features, matching statistics, and multimodal embeddings (both direct fusion and clustering‑based approaches) to alleviate cold‑start problems.

Consistency between recall, coarse‑ranking, and fine‑ranking is examined, introducing DNN‑based stacking and cascade models to align representations and reduce truncation loss.

Causal inference is applied in recall and coarse‑ranking by constructing pairwise samples of low‑popularity clicked items versus high‑popularity unclicked items, improving personalization for niche content.

Additional techniques include beam‑search sequence re‑ranking, graph databases and embeddings for user‑author interactions, and exploratory GNN‑based recommendation models.

The Q&A section addresses the importance of dwell time as a signal, handling model consistency during fail‑over, and the relationship between recall and ranking pipelines.