From Zero to One: Building 58.com Recruitment Personalized Recommendation System

The 58 recruitment platform serves millions of job seekers and employers daily, requiring a deep understanding of user intent and a robust recommendation pipeline to handle massive data, cold‑start issues, sparsity, and real‑time constraints.

Recommendation scenarios include job feed, category suggestions, and similar‑job recommendations for both C‑end seekers and B‑end companies, with content types such as job, tag, company, and resume recommendations.

User intent is captured through textual and behavioral signals, employing keyword/regex filters, phonetic matching, NER models (BiLSTM+CRF), and classification algorithms (fastText, CNN) to identify low‑quality or malicious users and to build a knowledge graph for richer profiling.

User profiling combines statistical rules, traditional classifiers, and sequence models (LSTM, GRU, Attention) to generate long‑ and short‑term interest vectors, supporting downstream recall and ranking.

Recall is performed via three complementary methods: contextual + user‑profile expansion, item‑based collaborative filtering with behavior weighting and time decay, and deep embedding‑based retrieval using FAISS for nearest‑neighbor search.

Ranking evolves through multi‑objective models predicting CTR, CVR, and ROR, with careful sample cleaning, feature engineering, and model monitoring; a feature pipeline automates sample generation, transformation, and combination.

The serving architecture enables automatic model updates, hot‑loading, and re‑ranking mechanisms to filter low‑quality content and improve effective connections.

Additional components include list‑page content control using NLG‑generated snippets and tag highlights, and an AB‑experiment configuration center for rapid iteration.

Future work focuses on multi‑task and reinforcement learning, expanding data sources for richer user portraits, and further personalization at scale.