CEU eTD Collection (2025); Zuroshvili, Elene: Fashion Intelligence through Multitask Learning: A Unified Model for Product Classification and Retrieval

CEU Electronic Theses and Dissertations, 2025
Author Zuroshvili, Elene
Title Fashion Intelligence through Multitask Learning: A Unified Model for Product Classification and Retrieval
Summary Online fashion platforms increasingly rely on automated systems to organize and retrieve prod- uct information. This thesis develops a multitask deep learning model that jointly performs category classification, attribute prediction, and image-based retrieval using the DeepFashion dataset. The model combines these tasks into a unified ResNet-50-based architecture with separate output heads, trained using cross-entropy and triplet loss. Five configurations are compared: single-task classifiers for category and attributes, a standalone retrieval model, a dual-head classification model, and a full multitask model incorporating all three objectives. While multitask learning introduces challenges in balancing competing losses, a staged training pipeline improves stability and leads to performance comparable to single-task baselines. Be- yond model performance, the thesis discusses how such systems can help reduce search frictions and improve product discoverability in online fashion markets.
Supervisor Posfai, Marton
Department Undergraduate Studies BA
Full texthttps://www.etd.ceu.edu/2025/zuroshvili_elene.pdf

Visit the CEU Library.

© 2007-2025, Central European University