Evolution of 58 Local Service Recommendation Algorithms: Scenarios, Tag & Post Recommendations, and Future Directions

The talk begins by describing the 58 Local Service recommendation scenarios, which span homepage category recommendations, secondary category pages, hot posts, hot services, and tag recommendations, all facing challenges like information homogeneity, complex user structures, mixed decision cycles, and a vast multi‑industry landscape.

To address these challenges, the system adopts fine‑grained scenario splitting and multi‑type recall strategies, including tag recall, vector recall (using DSSM with IndexFlatL2, IndexIVFlat, IndexIVFPQ), user‑intent recall (two‑level category prediction with BERT), and supplemental recall based on user‑selected categories.

Tag recommendation follows a classic recall‑ranking pipeline, evolving from simple statistical recall to TF‑IDF, word‑vector similarity, Bi‑LSTM, and finally the ATRank model, which leverages multi‑type embeddings and self‑attention to improve CTR by about 15% over the baseline.

Post recommendation architecture is divided into data, recall, ranking, and rendering layers. Recall combines multi‑channel methods, while ranking has progressed from linear models to tree models, deep models, and multi‑objective deep models (DeepFM and ESMM), achieving offline AUC around 0.78 and online CTR/CVR improvements.

The future outlook emphasizes strengthening foundational data (accurate tags, structured post information), enhancing core services (segmentation, synonym handling, embeddings), and balancing model accuracy with inference efficiency through techniques like model distillation, compression, and pruning.

A Q&A session covers practical aspects such as DSSM sample construction, CTR normalization, model deployment configurations, and troubleshooting deep learning models.