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

The presentation introduces the 58 recruitment commercial recommendation scenario, where a large C‑side job‑seeker audience and diverse B‑side job postings require efficient BC‑side matching to improve traffic monetization.

Scenario Overview : The APP’s top‑left "Find Job" entry drives 70% of traffic, split into non‑precise (hot, nearby) and precise (tag‑matched) recommendations. Job postings are abundant (6 × 10⁷ normal, 3 × 10⁵ commercial), with many fine‑grained categories, leading to cross‑category posting challenges.

System Architecture : A pipeline from multi‑channel recall to filtering and ranking is built using user and item data combined with state‑of‑the‑art machine‑learning algorithms. The architecture includes a strategy layer for content understanding, a recall‑filter‑ranking mechanism, and a final style‑optimization step before display.

Recall Methods :

Regional recall using DBSCAN clustering on user latitude/longitude to identify dense user regions; top‑clicked job titles in each region become hot‑post recommendations.

Behavior‑based recall employing the EGES model: sessions (3‑hour windows) are cleaned, a graph of clicked posts is built, random walks generate sequences, and a Word2Vec‑style training produces post embeddings enriched with side‑information (tags, category, salary). Vectors are unit‑normalized and concatenated with a 16‑bit city encoding (0 → -1) before cosine similarity retrieval.

Additional recall channels: content‑based (title Word2Vec), rule‑based (category expansion), and other heuristics.

Ranking :

Coarse Ranking aims to select 200‑300 candidates within 15 ms. It evolved from rule‑based to LR, then to a dual‑tower model (user tower and item tower decoupled, item vectors stored in Redis). Feature selection reduced dimensions from 450 to 150.

Knowledge distillation uses the fine‑ranking model as a teacher, combining soft targets (teacher predictions) with hard targets (click labels) to improve AUC from 0.68 to 0.70.

Fine Ranking addresses position bias and combines CTR and CVR. Models progressed from a baseline DIN to DIN‑bias (position sub‑network), then to W3DA (Wide‑Deep‑Attention) adding two wide sub‑networks for first‑ and second‑order feature crosses, finally to a multitask W3DA that jointly predicts CTR, CVR, and CTCVR (CTR × CVR) following the ESMM paradigm, achieving a 3.55 % lift in CTCVR while keeping CTR stable.

Evaluation & Q&A : Offline metrics focus on AUC; online performance is measured by CTR, CVR, ACP, and A/B tests. The Q&A covers DBSCAN evaluation, EGES embedding validation, negative sample selection for coarse ranking, hard vs. soft targets in distillation, and future plans such as edge deployment.

Future Work : Improve cold‑start recall by dynamically exploring user intent, adopt more expressive ranking models like DCNv2, and enhance traffic allocation strategies to foster a healthier commercial ecosystem.