Search List Ranking Efficiency Optimization Practices at 58.com

58.com, the largest domestic classified information platform, traditionally relied on time‑based sorting for list pages, which became inefficient as post volumes grew to millions; the need arose to return more relevant listings to users to boost click‑through and conversion.

The team built a comprehensive ranking framework that fuses timeliness, quality, business strategy, personalization, relevance, and deduplication across coarse, fine, and re‑ranking stages, and applies it to all product lines.

The optimization roadmap consists of four stages: feedback strategy, basic model, feature upgrade, and model upgrade, each addressing increasingly sophisticated challenges.

In the feedback strategy stage, historical click‑through and conversion rates are used as ranking factors, with solutions for cold‑start (Bayesian smoothing, UCB, Thompson sampling), position bias (COEC correction), and time decay (exponential decay, later abandoned for simplicity).

The basic model stage focuses on data pipelines—log merging and cleaning, feature engineering, and model training—using LR and GBDT models, extensive feature selection (filtering, FCBF) and processing (discretization, one‑hot, normalization), and rigorous offline/online evaluation and consistency checks.

Feature upgrades introduce timeliness improvements, Cartesian and matching feature combinations, and text/image embeddings (word‑vector averaging, deep residual network features), substantially enhancing model performance.

Model upgrades add FM for automatic feature interactions, result and feature fusion techniques, and deep learning models (DNN, FNN, Wide&Deep, DeepFM), with Wide&Deep selected for production, yielding measurable AUC gains.

Overall, the iterative approach delivered 20%+ conversion improvements in rental listings, 30%+ ECPM gains in premium listings, and similar lifts in second‑hand car listings; future work will continue to enhance features, explore deeper models, and solidify the end‑to‑end efficiency‑optimization platform.