Ranking Strategy Optimization Practices for Commercial Traffic at 58.com

The commercial strategy team at 58.com describes how they improved the ranking of commercial traffic across multiple business lines (local services, recruitment, real‑estate, used cars, etc.) by upgrading data pipelines, features, models, and the ranking mechanism.

Background : In the listing mode, large candidate sets make ranking crucial for monetization efficiency and user experience. The original baseline model consisted of four parts: basic data flow, feature engineering, model computation (FTRL, GBDT, wide&deep), and a single‑factor ranking (CTR, CVR, eCPM).

Optimization Path :

1. Data‑flow upgrade : Two real‑time pipelines were built—one using Spark Streaming for real‑time impression‑click stitching and another using Flink for real‑time metric calculation. This reduced feature diff from ~5% to <1% and cut sample generation latency to about one minute.

2. Feature upgrade : ID‑type features (e.g., car model, brand) were transformed with embedding and attention mechanisms. User‑interest vectors were generated by embedding discrete attributes, weighting them by interaction counts, and pooling to a fixed‑length vector. Cross‑features between user behavior and post attributes were also embedded.

3. Model upgrade :

a) Online training : Real‑time samples enable online training of an FTRL model, with model snapshots dumped every 10 minutes and hot‑loaded in production.

b) Multi‑task learning (ESMM) : A shared‑embedding ESMM model jointly predicts CTR and post‑click conversion (PCTCVR) on the full impression space, alleviating sample‑selection bias and data sparsity.

c) Relevance model : Coarse‑ranking uses cosine similarity between user and post vectors, combined with weighted strong and weak intent scores.

4. Ranking mechanism upgrade : The new framework supports layered sorting, weighted fusion, custom fusion (e.g., multiplication), and nested layers, allowing multiple factors (CTR, CVR, relevance, business rules) to be combined flexibly.

Results and Outlook : After the four upgrades, the multi‑factor ranking system shows significant gains in monetization and connection efficiency across business lines. Future work includes further data‑cleaning models, automated feature engineering, multi‑model fusion, model compression, and automated factor exploration for multi‑objective offline evaluation.

References :

MaX, Zhao L, Huang G, et al. Entire space multi‑task model: An effective approach for estimating post‑click conversion rate. SIGIR 2018.

Huang P S, He X, Gao J, et al. Learning deep structured semantic models for web search using click‑through data. CIKM 2013.

Barkan O, Koenigstein N. Item2vec: Neural item embedding for collaborative filtering. IEEE MLSP 2016.

Author : Liu Lixi, Senior Algorithm Engineer, Strategy Technology Team, Commercial Product Technology Department, 58.com, focusing on commercial traffic monetization and ranking optimization.