Idea
Model improving e-commerce search relevance by learning query-aware discrete semantic identifiers for precise product ranking.
Research Paper
Core Innovation
This paper introduces DSIRM, which explicitly models discrete relevance features by bridging queries and items through contrastive quantization and generative LLMs. Unlike prior unsupervised SID methods, it injects query-item interaction supervision to learn relevance-aware semantic partitions, enabling better query-dependent ranking and handling of ambiguous queries.
Why It Matters
E-commerce platforms struggle to capture fine-grained product attributes and query intent, limiting search relevance and user satisfaction. DSIRM enhances ranking accuracy by integrating query-item interactions into discrete semantic identifiers, improving conversion rates and user engagement. This scalable approach addresses tail queries and intent ambiguity, critical for large-scale online retail.
Market Size (TAM)
$20–50B TAM for e-commerce search relevance platforms; $2–10B SAM from large online marketplaces and retailers. Driven by growth in online shopping and demand for personalized search experiences.
Potential Customers & Pain Points
- E-commerce platforms – Poor search relevance and attribute distinction
- Online marketplaces – Difficulty handling tail queries and ambiguous intents
- Retailers – Low conversion rates due to imprecise product ranking
Business Model
Licensing the DSIRM model as a SaaS API or on-premise solution to e-commerce platforms and marketplaces, with tiered pricing based on query volume and feature usage.
Competitive Landscape
- Amazon A9
- Google Shopping Search
- Alibaba Search
- Coveo
- Bloomreach
Implementation Challenges
- Integration complexity with existing search infrastructure
- Dependence on large-scale query-item interaction data
- Handling diverse and evolving product catalogs
Validation Strategy
- Pilot deployment on mid-sized e-commerce platforms to measure CTR and conversion lift
- A/B testing against existing search relevance models in production
- Collecting user feedback and query logs to refine query-item interaction modeling
Research Paper Overview
DSIRM: Learning Query-Bridged Discrete Semantic Identifiers for E-commerce Relevance Modeling
Summary
This paper presents DSIRM, a model that improves e-commerce search relevance by learning discrete semantic identifiers with query interaction supervision, enhancing fine-grained attribute distinctions and query-dependent ranking. It combines contrastive quantization on items and generative LLMs on queries to predict item identifiers, achieving significant offline and online performance gains on Tmall's platform.