Visual Uniqueness: An Unsupervised Contrast Learning Approach

Simple item page
dc.contributor.author Feng, Xiaohang
dc.contributor.author Li, Charis
dc.contributor.author Zhang, Shunyuan
dc.date.accessioned 2023-12-26T18:38:40Z
dc.date.available 2023-12-26T18:38:40Z
dc.date.issued 2024-01-03
dc.identifier.isbn 978-0-9981331-7-1
dc.identifier.other 25b476ca-562f-4782-adf5-45e09277b7c7
dc.identifier.uri https://hdl.handle.net/10125/106686
dc.language.iso eng
dc.relation.ispartof Proceedings of the 57th Hawaii International Conference on System Sciences
dc.rights Attribution-NonCommercial-NoDerivatives 4.0 International
dc.rights.uri https://creativecommons.org/licenses/by-nc-nd/4.0/
dc.subject Data Analytics, Data Mining, and Machine Learning for Social Media
dc.subject airbnb
dc.subject contrastive learning
dc.subject controlled experiment
dc.subject eye-tracking
dc.subject image analytics
dc.subject visual uniqueness
dc.subject xai (explainable ai)
dc.title Visual Uniqueness: An Unsupervised Contrast Learning Approach
dc.type Conference Paper
dc.type.dcmi Text
dcterms.abstract This paper develops an unsupervised machine learning model that scores a product image on its visual uniqueness. Based on large-scale images of Airbnb properties in New York City, our model used contrastive loss and random data augmentation to compute the visual uniqueness of a property image automatically. The model achieves 88.10% accuracy on a hold-out set. We identified key image features that make a room unique. Leveraging the advanced explainable AI techniques to generate interpretable uniqueness heatmaps, we found certain decorations (e.g., pillows, paintings) may help enhance room uniqueness. Next, we validated the model against human perceptions via two lab studies and an eye-tracking controlled experiment: both the model-predicted uniqueness and key image features are consistent with human judgment. We discussed discriminative validity between uniqueness and aesthetics. This research offers important managerial implications for individual hosts to optimize the visual presentation to stand out in the crowded market.
dcterms.extent 11 pages
prism.startingpage 2505
Files
Original bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
0244.pdf
Size:
1.48 MB
Format:
Adobe Portable Document Format
Description:
Download
Collections
Data Analytics, Data Mining, and Machine Learning for Social Media