Technical Evolution of the “Guess You Want” Recommendation Module in 58 Local Services

Background – 58 Local Services provides a platform for users to find everyday services, but many search landing pages suffer from empty results. To improve conversion, a “Guess You Want” module was added to recommend relevant posts when user queries return few or no results.

Overall Framework – The system consists of four layers: data, recall, ranking, and rendering. The data layer aggregates user logs and business data; the recall layer retrieves candidate posts via multiple strategies; the ranking layer orders candidates using model scores and business rules; the rendering layer displays the final results.

Recall Layer – A multi‑path recall approach is employed: (1) intent‑based recall using hierarchical BERT classifiers to predict first‑level and second‑level categories; (2) vector recall using a DSSM model whose embeddings are indexed with FAISS; (3) tag‑based recall that expands queries with high‑quality business tags; and (4) a supplemental hot‑post recall for cases where other strategies return nothing.

Intent‑Based Recall – User queries are first classified into coarse categories (≈10 first‑level categories) and then refined to finer categories (≈200 second‑level categories) using a two‑stage BERT model. Training data are derived from click logs, post titles, and combined query‑title pairs, with extensive data augmentation and noise filtering.

Vector Recall – Offline, a DSSM model learns embeddings for queries and post titles; these embeddings are stored in a FAISS index (IndexIVFFlat). Online, a query is embedded and the nearest vectors are retrieved as candidate posts.

Tag Recall – The rich tag system of 58 is leveraged to extract relevant tags from the query, which are then used by the main search service to fetch additional matching posts.

Supplementary Recall – When all other recall paths fail, popular posts are selected based on historical call‑to‑view ratios, filtered by city and category.

Ranking Layer – Features from users, posts, merchants, and context are combined. Initial models used XGBoost, but due to limited feature interaction, deeper models were explored.

Model Evolution – DeepFM was introduced to capture both low‑order (FM) and high‑order (DNN) interactions. To handle multiple objectives (CTR and CVR), a Multi‑Gate Mixture‑of‑Experts (MMoE) model was built. ESMM was later adopted to jointly learn CTR and post‑click conversion (CTCVR) across the entire sample space. Finally, a hybrid DeepFM+ESMM architecture was prototyped to merge low‑order, high‑order, and multi‑task learning.

Conclusion – The “Guess You Want” system effectively reduces no‑result rates and boosts user conversion by integrating diverse recall methods and progressively more sophisticated ranking models, while future work will continue to tailor the recommendation pipeline to business needs.