Embedding Techniques for Advertising Recall and Ranking in a Second-Hand Car Platform

The commercial strategy team presents a comprehensive study of embedding technologies applied to a local‑life second‑hand car platform, aiming to improve both advertising recall and ranking while balancing user experience and revenue.

Mainstream Embedding Techniques – The article introduces three primary categories: Skip‑Gram models (e.g., word2vec, item2vec), graph‑embedding (e.g., DeepWalk), and NN‑embedding using multi‑layer neural networks. Each method’s strengths, weaknesses, and suitability for different scenarios are discussed.

Second‑Hand Car Application Scenario – Characteristics such as complex user behavior, highly structured vehicle information, and diverse commercial products (CPC, CPT, CPA) create challenges for matching efficiency. Embedding is employed to enhance matching across these varied contexts.

Embedding in Advertising Recall – The recall pipeline consists of AB‑test traffic splitting, recommendation algorithms, result fusion, and filtering. Both collaborative‑filtering and embedding‑based vector recall are deployed. A Skip‑Gram‑based item2vec model and a NN‑embedding model are trained on user sessions, with negative sampling and re‑sampling of conversion events. Faiss is used for large‑scale vector search, with daily index refreshes to handle item turnover.

Embedding in Advertising Ranking – User behavior is split into explicit (filter/search) and implicit (click/phone) intents. Explicit intents are modeled with a bag‑of‑words approach, while implicit intents use embedding vectors. A dual‑tower DSSM architecture (user tower and item tower) with shared embeddings generates user and item vectors. Loss functions are adjusted to cosine similarity + cross‑entropy to incorporate both positive and negative samples.

Coarse‑Ranking Integration – After success in fine‑ranking, the relevance factor is introduced into coarse‑ranking to improve candidate quality. Optimizations include using recent user actions, pre‑training item vectors with item2vec, changing the loss to cross‑entropy, and adding diversity factors.

Evaluation – The system is evaluated through manual relevance checks, PCA visualizations of item embeddings, user embedding relevance tests, and online A/B experiments. Results show a ~10% lift in conversion rate for recall, a 12.36% NDCG improvement in ranking, and a 5% increase in user connection rate in coarse‑ranking.

Conclusion and Outlook – Embedding techniques have been successfully deployed across recall, coarse‑ranking, and fine‑ranking, delivering measurable business gains. Future work includes time‑aware and graph‑based embeddings to capture richer interactions, emphasizing that “everything can be embedded.”