Taxonomy-based Negative Sampling In Personalized Semantic Search for E-commerce

Uthman Jinadu, Siawpeng Er, Le Yu, Chen Liang, Bingxin Li, Yi Ding, Aleksandar VelkoskiarXiv 2025 · DOI · arXiv

Introduction

E-commerce search presents unique challenges compared to general web search. Queries are often short, highly ambiguous, and deeply dependent on individual user intent. Traditional lexical search engines, which rely on exact keyword matching (e.g., BM25), frequently fail to bridge the semantic gap between a user's query and relevant products. While modern neural "two-tower" models offer a solution by embedding queries and items into a shared vector space, they face significant hurdles in training efficiency and personalization.

Researchers from Georgia State University and The Home Depot have developed a novel semantic search framework that addresses these issues. By integrating user-level personalization directly into the retrieval stage and employing a novel Taxonomy-Based Hard-Negative Sampling (TB-HNS) strategy, the framework achieves substantial improvements in product discovery and conversion rates.

The overall system architecture integrating the Semantic Engine Model into the search pipeline.

System Architecture and Methodology

The proposed framework integrates a semantic engine into a multi-stage search pipeline. As shown in the system overview, the process involves de-duplication, filtering, and multiple ranking stages (Pre-Rank, Rank, and Re-rank) before displaying items to the user. The core of this innovation lies in the Semantic Engine Model, which operates both online for real-time query embedding and offline for catalog indexing.

Personalized Semantic Retrieval

Most existing systems treat retrieval and personalization as distinct, sequential steps. This framework introduces a "Customer Tower" that fuses the search query with user-specific features, such as demographics and past purchase history, using a dense layer. This creates a personalized query embedding ($q_c$) before the search interacts with the product index.

The personalized two-tower architecture featuring a Customer Tower and an Item Tower to generate similarity scores.

This architecture utilizes Sentence-BERT (SBERT) to process both the query and item attributes (Title, Brand, Color, Category). By calculating the similarity score between the personalized customer embedding and the item embedding, the model can surface products that align with both the literal query and the user's historical preferences.

Taxonomy-Based Hard-Negative Sampling (TB-HNS)

A critical challenge in training these models is negative sampling—providing the model with examples of what not to retrieve. Random sampling often produces "easy" negatives (e.g., a refrigerator for a "hammer" query), while traditional hard-negative mining is computationally expensive.

TB-HNS leverages the existing product hierarchy to find "near-miss" negatives. For a given positive product, the algorithm identifies its parent category and samples "sibling" products from the same branch. This forces the model to learn fine-grained distinctions (such as brand or finish) within a relevant context without the overhead of global index scans.

Key Results and Impact

The system was evaluated using a massive internal dataset from The Home Depot and the public Amazon ESCI dataset.

  • Offline Performance: TB-HNS achieved a 38–48% relative gain in Recall@K compared to industry-standard baselines like ANCE. For specific queries, personalization improved Recall@8 by 98.68%.
  • Live A/B Testing: In production, the model delivered a 2.70% increase in Conversion Rate (CR) and a 2.04% increase in Add-to-Cart (ATC) rates.
  • Operational Efficiency: The system maintained a P95 latency of ~50ms for embedding generation, meeting the strict requirements of high-traffic e-commerce environments.

Conclusion

By combining taxonomy-aware training with early-stage personalization, this framework significantly enhances the relevance of e-commerce search. Future work will explore multi-modal embeddings, incorporating product images to further refine the semantic understanding of the catalog.

Summary updated

BibTeX
@misc{jinadu2025taxonomy,
  author = {Uthman Jinadu and Siawpeng Er and Le Yu and Chen Liang and Bingxin Li and Yi Ding and Aleksandar Velkoski},
  title = {{Taxonomy-based Negative Sampling In Personalized Semantic Search for E-commerce}},
  year = {2025},
  eprint = {2511.00694},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.00694},
}
Figure