Embedding Techniques in 58 Commercial Search and Recommendation: Practices and Insights

Speaker Su Yonghao, senior algorithm engineer at 58.com, introduces the commercial traffic scenarios of 58, including search, recommendation, and filtering, and explains how embedding is used to capture user intent and business promotion capabilities.

58 Commercial Traffic Scenarios : Search has explicit user intent expressed by queries, recommendation deals with implicit intent inferred from user behavior, and filtering imposes explicit constraints while still focusing on user interest.

Mainstream Embedding Algorithms : Four categories are presented – matrix‑factorization (e.g., SVD), content‑based embeddings (static: word2vec, fasttext; dynamic: ELMo, GPT, BERT), item‑sequence embeddings (item2vec), and graph embeddings (shallow models like DeepWalk, Node2vec, EGES; deep models like GCN, GraphSAGE).

Static representations such as word2vec generate a single vector per word; they are efficient but cannot handle polysemy. Word2vec uses CBOW or Skip‑gram architectures with hierarchical softmax or negative sampling for optimization.

Dynamic representations like ELMo (bidirectional LSTM) and BERT (stacked transformer encoders with masked language modeling and next‑sentence prediction) produce context‑dependent embeddings, overcoming the limitations of static methods.

Sequence embeddings (e.g., item2vec, Hema Embedding) incorporate item attributes (product ID, brand ID, category ID) to alleviate cold‑start problems. Graph embedding EGES builds a product‑relation graph from user behavior and applies DeepWalk with side‑information‑aware skip‑gram.

Industry examples include Airbnb Embedding (clustered item embeddings with improved negative sampling) and YoutubeNet (concatenating user continuous, sequential, and discrete features into a DNN, using the final softmax layer as video embeddings).

Embedding in 58 Commercial Search : The search pipeline consists of query understanding (pre‑processing, tokenization, query rewriting, term analysis, intent recognition), retrieval & recall (text, tag, semantic channels), matching (semantic, textual, tag similarity), and ranking (balancing relevance, CTR, CVR, and business metrics). Query expansion uses ESIM (a bi‑LSTM interaction model) to generate related queries, later refined with a fine‑tuned BERT model that ranks candidate terms based on relevance scores.

To handle category‑specific intent, class information is appended to the premise query before interaction, improving disambiguation.

Embedding in 58 Commercial Recommendation : The recommendation system comprises recall and ranking layers. Recall aggregates multiple channels (rules, traditional models, vector‑based). Ranking employs multi‑factor models estimating CTR, CVR, and relevance, with relevance modeling enhanced by embedding.

Weak intent (click, call, chat) is modeled with a dual‑tower architecture; strong intent (search, filter) uses logistic regression with time decay. User and item towers share an embedding lookup table; item attributes (including geographic information) are encoded to address cold‑start and locality constraints.

For coarse‑ranking, a hybrid of DSSM, Airbnb Embedding, and YoutubeNet ideas is used: item towers encode attributes, user towers concatenate behavior sequence embeddings with user attribute embeddings, and historical statistical factors are fused before the final DNN output. The system updates item vectors offline (weekly) and user vectors in real‑time.

Online data flow asynchronously generates user and item vectors, supporting relevance estimation in coarse‑ranking, fine‑ranking, and recall stages.

Evaluation shows clear separation of vectors on brand and price dimensions via PCA, and significant improvements in relevance metrics for both coarse and fine ranking after deployment.

Q&A Highlights :

Embedding effectiveness is evaluated through visualization, business‑specific recall analysis, and top‑similar post inspection.

User‑item vectors are constrained to the same space to enable similarity computation.

Graph embeddings (e.g., EGES) are used in recall, though not the focus of this talk.

Virtual taxonomy is built either by expert curation or data‑driven clustering of keywords.

Attention mechanisms were tested for user sequence aggregation, but recent‑visit average‑pooling performed best.

Relevance modeling is applied in both coarse and fine ranking, with different emphasis on relevance vs. revenue.

The presentation concludes with thanks and a reminder to like, share, and follow.