Home // GPTMB 2024, The First International Conference on Generative Pre-trained Transformer Models and Beyond // View article

Using Bi-Directional Instance-Based Compatibility Prediction for Outfit Recommendation

Authors:
Tzung-Pei Hong
Yun-Pei Chao
Jiann-Shu Lee
Ja-Hwung Su

Keywords: outfit recommendation; fashion compatibility; Bi-LSTM; deep learning

Abstract:
Existing fashion recommendation studies focus primarily on recommending individual items. However, this paradigm cannot cater to user needs on fashionable outfit. To obtain a fashionable and well-coordinated outfit, outfit recommendation focuses not only on one item but on all items in an outfit. Such fashion recommendation outputs multiple images of items to constitute a whole outfit. To this end, this paper proposes a novel outfit recommendation method named Bi-directional Instance-based Compatibility Prediction (BICP) suggesting suitable revised outfits based on the outfit inputs of users. In this method, the conditional Bi-directional Long Short-Term Memory (Bi-LSTM) mechanism is used as a backbone to generate the embedding representation of fashion items. To approximate the best outfit, a new metric called I2I-cos (Instance-to-Instance) cosine similarity is also proposed for outfit compatibility calculation. Finally, we made distribution diagrams indicating the outfits recommended by the proposed approaches better align with people's aesthetics and preferences.

Pages: 36 to 40

Copyright: Copyright (c) IARIA, 2024

Publication date: June 30, 2024

Published in: conference

ISBN: 978-1-68558-182-4

Location: Porto, Portugal

Dates: from June 30, 2024 to July 4, 2024