Graph-Based Learning for Taxonomy Optimization

Taxonomies on e-commerce platforms naturally form directed acyclic graphs that organize millions of products into hierarchical structures critical for buyer navigation and product discovery. However over time these taxonomies tend to accumulate fragmentation, inconsistent granularity, and misaligned boundaries, leading to buyer confusion and frustration and decreased conversion rates, prompting increased manual curation efforts on an already overloaded taxonomy. To alleviate these issues in a scalable way, we present a graph-native framework for taxonomy reduction and optimization - a setting that, while crucial for large, heavily-fragmented taxonomies, remains relatively underexplored. Modeling the taxonomy as a directed graph, categories are enriched with LLM-generated textual descriptions and encoded via Sentence Transformers. We then employ a Graph Neural Network architecture trained with margin-based triplet loss to learn semantically and structurally informed embeddings. Cosine similarity on the learned space produces actionable proposals that merge semantically redundant categories while preserving distinctions relevant to buyer intent. A panel of LLM judges evaluates each proposal; in subsequent auditing, human judges deemed 88% of the LLM-approved merges correct. Offline evaluations show a marked increase in the number of items retrieved for popular queries after applying the proposed merges, while retaining, and in early ranks marginally improving, top-item relevance. Overall, the approach offers a scalable, domain-adaptive method for restructuring large e-commerce taxonomies, reducing dependence on resource-intensive manual processes while maintaining strict precision requirements.