The Evolution of AutoHome Recommendation System Ranking Algorithms

AutoHome's recommendation system, operating for nearly five years, serves billions of resources across articles, videos, images, and more, aiming to match users with personalized content through three sub‑goals: understanding users, characterizing resources, and optimal matching.

The system pipeline consists of four stages—resource collection, candidate recall, ranking, and final top‑N output—implemented via modules such as a resource pool, tag generation, indexing, filtering, recall, user profiling, ranking models, feature handling, and operational controls.

Ranking model development progressed from Logistic Regression (LR) to XGBoost, FM, Wide&Deep, DeepFM, DCN, and sequence models (LSTM, GRU), with each iteration delivering incremental CTR improvements. DeepFM became the baseline model, while online learning now updates models at minute‑level intervals using real‑time label and feature joins.

Online ranking services expose an API that accepts device ID, item IDs, and model specifications, returning scored items. The service relies on a feature service that provides both offline (historical) and real‑time user/item features, processed through anomaly handling, normalization, and equal‑frequency bucketing.

Feature engineering includes user attributes, behavior statistics, item metadata, cross features, and multimodal embeddings (text BERT, image/video embeddings, graph embeddings). Feature production combines offline batch data with second‑level real‑time updates.

Future directions target multi‑objective optimization (CTR, dwell time, interaction), advanced model architectures (Transformers, AutoML, reinforcement learning), richer user interest embeddings, and deeper multimodal information fusion.