Summary of 58 Group Technical Salon: Recommendation System Architecture and Search Ranking Algorithm Practices

On May 17, 2019, the 58 Group Technical Salon (Session 11) titled "Recommendation Architecture and Algorithm Practice" was held at the Beijing headquarters, featuring speakers from leading internet companies who shared experiences on recommendation system architecture and search ranking algorithms.

1. 58 Recommendation System Architecture Practice

The recommendation system has been refactored into a stable, high‑performance, low‑coupling microservice architecture that supports rapid policy iteration and is deployed across all 58 recommendation scenarios. The system consists of three layers:

Data Layer: Stores business data (user and post data) and user behavior logs. Data is persisted both in batch on HDFS for offline analysis and streamed to Kafka for real‑time computation.

Strategy Layer: Implements recall and ranking. Recall combines collaborative filtering, matrix factorization, DNN, regional hot items, interest recall, and association rules. Ranking uses machine‑learning models such as LR, FM, GBDT, fused models (GBDT+LR, GBDT+FM) and DNN, with offline processing via MapReduce and Hive, multidimensional analysis with Kylin, large‑scale training with Spark/DMLC, and deep learning with TensorFlow on the 58 AI platform.

Application Layer: Exposes RPC and HTTP interfaces for external services, supporting scenarios like "You May Like", detail page relevance, search no‑result recommendation, feed stream, personalized, and hot recommendations, typically delivering top‑N recommendation lists.

2. 58 Search Ranking Algorithm Practice

The self‑developed search machine‑learning platform provides highly configurable pipelines for sample selection, feature engineering, and model training, addressing the challenges of maintaining multiple business‑specific processes.

Sample Sampling: Solves class imbalance with basic filters (AllPassFilter, AllNotPassFilter, ClassFilter, ELFilter, PosFilter) and combinatorial operators (AndFilter, OrFilter, NotFilter).

Feature Engineering: Includes personalized features (user‑post interactions), continuous features (equal‑frequency, equal‑width, custom), and discretization/normalization (one‑hot encoding).

Model Training: Supports basic models (LR, FM, GBDT, XGBoost) on Spark and fusion models that combine multiple base models, model predictions as features, FM factors, or XGBoost tree indices.

The platform offers a web UI for configuration and model management.

3. Summary

The salon facilitated knowledge exchange on recommendation system architecture, search ranking algorithms, and related technical challenges, providing valuable insights for future system improvements and efficiency optimizations.