Recruitment Recommendation System: Ranking Framework, Model Evolution, and Feature Engineering

58.com is China’s largest classified information platform, offering services such as recruitment, housing, used cars, and directories; recruitment is a core business serving millions of job seekers and employers.

The platform faces low matching efficiency and a lack of personalization, with user demands varying across industry, season, region, and demographic dimensions.

Typical recommendation scenarios include homepage recommendations, tag‑based suggestions (recent interests, bottom tags), area‑based recommendations (nearby jobs, urgent hires), detail‑page recommendations ("also viewed", nearby postings, similar popular jobs), and chat‑based job searching.

The ranking framework consists of three layers: a coarse recall stage that filters candidates based on job freshness, B‑side activity, and quality factors; a fine‑grained ranking stage that uses machine‑learning models to predict user interest and employer satisfaction; and a re‑ranking stage that balances resource allocation, user diversification, and commercial factors.

Optimization targets aim to maximize Gross Connection Volume (GCV) by decomposing it into three sub‑models: click‑through rate (CTR), conversion rate (CVR), and feedback rate (FBR).

Model evolution progressed through five stages: handcrafted rules, large‑scale discrete Logistic Regression (LR) with FTRL optimization, feature engineering, non‑linear models (FM, XGBoost), and deep learning models.

In the rule‑based stage, domain experts designed weight‑adjusted rules (e.g., matching job category with user preference) and iteratively refined them via A/B testing.

The large‑scale discrete LR stage treated the prediction problem as a supervised binary classification task, discretizing continuous features and employing algorithms such as Spark’s mini‑batch SGD, LBFGS, and later Google’s FTRL to achieve sparse, efficient models suitable for billions of sparse features.

Feature engineering covered four categories: basic features (all factors influencing recruitment outcomes), statistical features (multi‑window statistics, confidence, smoothing), combinatorial features (crossed features based on business logic, e.g., gender × job type), and text‑label features derived from natural‑language processing to capture richer job and user semantics.

Non‑linear models introduced factorization machines (FM) to automatically learn second‑order feature interactions, and XGBoost trees to capture higher‑order combinations, with the XGBoost output one‑hot encoded and concatenated with original sparse features for a stacked LR model.

Deep learning models, such as Wide&Deep, DeepFM, DeepCross, and xDeepFM, were explored to reduce feature‑engineering effort and capture high‑order implicit interactions; DeepFM combines a wide linear part with a deep neural network sharing embeddings, achieving a 1 % AUC lift over the LR+XGBoost baseline in CVR prediction.

The summary emphasizes that improving the bilateral connection efficiency between job seekers and employers is central to 58.com’s recruitment ranking; the task is decomposed into CTR, CVR, and FBR sub‑models, with a historical evolution from rule‑based systems to LR, FM, XGBoost, and deep learning approaches.

Future work will focus on real‑time streaming learning for minute‑level model updates, multi‑task learning across business objectives, attention‑based user‑behavior modeling, and listwise ranking methods.

References to academic papers and open‑source projects are listed, followed by a brief team introduction.