Evolution of 58 Local Service Recommendation Algorithms and Future Directions

The presentation introduces the 58 local service recommendation platform, highlighting the challenges of heterogeneous information, complex user structures, varied decision cycles, and multi‑industry, multi‑scenario, multi‑objective requirements.

Recommendation Scenarios and Characteristics : Four main app scenarios are described—homepage category recommendation, secondary category and hot post/service/tag recommendation, and detail‑page related services. The scenarios suffer from information homogeneity, complex user groups (unlogged, new, low‑activity), and mixed short/long decision cycles across over 200 industries.

Tag Recommendation Evolution : The pipeline follows a recall‑then‑ranking framework. Multi‑path recall includes context (search terms, clicked tags), statistics (hot clicks/searches), and long/short‑term user behavior. Model progression moved from simple statistical recall to TF‑IDF, word‑vector similarity, and finally behavior‑driven models such as Bi‑LSTM and ATRank, achieving up to 15% CTR lift over the baseline.

Post Recommendation Evolution : The architecture consists of data, recall, ranking, and rendering layers. Recall employs multi‑path strategies: user‑intent recall (two‑stage category prediction using BERT), tag recall, vector recall (DSSM with IndexFlatL2, IndexIVFlat, IndexIVFPQ), and supplemental recall based on user‑selected categories. Ranking progressed from linear models to tree models, deep models, and multi‑objective deep models, incorporating features from users, posts, merchants, and context. DeepFM combines FM and DNN for high‑order feature interactions, while ESMM jointly optimizes CVR and CTR, delivering higher offline AUC and online conversion metrics.

Future Outlook : Planned improvements include enhancing tag system accuracy and post structural information, strengthening foundational services (segmentation, synonym handling, embeddings), and balancing model accuracy with inference latency through techniques like model distillation, compression, and architecture pruning.

Q&A Highlights :

DSSM uses recent three‑month logs for weekly incremental training; positives are query‑post pairs leading to calls, negatives are random.

CTR features are normalized by capping extreme values and scaling to [0,1].

Different scenarios use different models; DeepFM/ESMM for "you may like" while simpler models serve category recommendations.

Evaluation of recall is performed by incremental addition of strategies and manual relevance checks.

Training runs on P40 GPUs, with distributed training for large datasets.

When DeepFM fails to converge, investigate code bugs and feature relevance; negative sampling is typically set to a 1:3 positive‑negative ratio.