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

The article introduces 58.com’s recruitment business and explains why personalized recommendation is crucial for matching millions of job seekers with employers, outlining three main sections: business overview, recommendation practice, and insights with future plans.

It describes the recommendation scenarios for both C‑end job seekers and B‑end enterprises, including job, tag, company, and resume recommendations, and highlights challenges such as massive data computation, cold‑start, sparsity, real‑time behavior, and resource allocation.

User understanding is achieved by analyzing both textual and behavioral signals to build a knowledge graph and detailed user profiles, employing NER (BiLSTM+CRF) and various feature extraction methods to identify genuine intent and filter out low‑quality or malicious users.

The recall module evolves from context‑based precise recall, improved collaborative filtering tailored to recruitment, and deep embedding‑based recall that vectorizes jobs and users, leveraging techniques like word2vec, FAISS, and multi‑task learning.

Ranking progresses through multi‑objective models (CTR, CVR, ROR), extensive feature pipelines, and a serving framework that supports automatic model updates, monitoring, and AB testing, while a re‑ranking mechanism adds quality and resource‑efficiency controls.

Additional components include list‑page content control using NLG‑generated descriptions and tag highlights, an AB experiment configuration center for rapid iteration, and an overall technical architecture that integrates offline data warehouses, knowledge graphs, user profiling, and online recommendation engines.

The article concludes with lessons learned—emphasizing deep business‑algorithm understanding, robust feature engineering, and tooling—and outlines future work such as multi‑task learning, reinforcement learning, and expanding data sources for richer user portraits.