Semantic analysis and clustering of customer search queries for e-commerce offer optimization
Year: 2026
Type: dissertation
Abstract: Large e-commerce platforms receive thousands of short, noisy, and highly variable search queries every day. Misspellings, abbreviations, product codes, and ambiguous terms limit the effectiveness of traditional keyword-based handling and contribute to zero-result searches, degraded user experience, and missed business opportunities. This thesis proposes a semantic pipeline to represent customer search queries in a shared embedding space and to organize them into an interpretable hierarchical structure that can support both improved handling of zero-result searches through similarity-based retrieval and assortment-related insights by revealing recurrent demand patterns that are not well covered by the current catalog. Using MediaMarkt ES search logs, we build a stratified sample queries preserving the zero-result proportion. Several sentence-embedding models are compared via distance-matrix agreement against a lexical baseline and two candidates are selected for expert A/B evaluation through a Google Forms questionnaire. The selected embeddings are then clustered to produce a deployable taxonomy, comparing hierarchical agglomerative clustering, bisecting k-means and HDBSCAN. Finally, cluster labels are generated to improve readability and the hierarchy is delivered as an interactive dendrogram with zero-result queries highlighted. Thereafter, experts are requested again for choosing among the best output.
Author: Barraquer Torres, Bruna
Institution: Universitat Politècnica de Catalunya
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7154797535