Commercial Recommendation System for 58 Recruitment: Architecture, Recall, and Ranking Techniques

The presentation introduces the commercial recommendation scenario of the 58 recruitment app, where the majority of traffic comes from recommendation slots and both B‑side advertisers and C‑side job seekers have diverse needs.

It outlines the overall system architecture, which integrates content understanding, multi‑stage recall, filtering, and ranking, and emphasizes the importance of aligning B‑side revenue goals with C‑side relevance.

Recall stage: Two main approaches are described. (1) Region‑based recall uses DBSCAN clustering on user GPS data to identify dense user areas and then selects hot job posts within those regions. (2) Behavior‑based recall employs the EGES graph‑embedding model: user sessions (three‑hour windows) are cleaned, a session‑level graph of clicked posts is built, random walks generate sequences, and a Word2Vec‑style training produces post embeddings, which are combined with a 16‑bit city encoding (0→‑1) for final vector similarity search.

Additional recall methods include content‑based recall using title Word2Vec, rule‑based recall based on category expansion, and other heuristics.

Ranking stage: It is divided into coarse and fine ranking. Coarse ranking aims to prune thousands of candidates to a few hundred within ~15 ms. Models evolved from rule‑based to logistic regression, then to a dual‑tower architecture, and finally to knowledge‑distillation where a complex fine‑ranking model provides soft targets. Feature selection reduced dimensions from 450 to 150.

Fine ranking addresses position bias and improves relevance. Models progressed from a simple position feature to a DIN‑bias network (position sub‑network + Deep Interest Network), then to W3DA (wide sub‑networks for first‑order and second‑order feature crosses combined with DIN), and finally to MultiTask‑W3DA that jointly models CTR and CTCVR using an ESMM‑style approach, achieving higher conversion while maintaining CTR.

The talk concludes with future work: more sophisticated cold‑start intent exploration, adoption of stronger models such as DCNv2 for ranking, and continued optimization of recommendation mechanisms and traffic allocation to foster a healthier commercial ecosystem.

A Q&A session follows, covering evaluation metrics for DBSCAN clustering, EGES embedding quality, negative‑sample selection for the dual‑tower model, hard vs. soft targets in knowledge distillation, and deployment considerations.