Chic or Social: Visual Popularity Analysis in Online Fashion Networks Kota Yamaguchi 1 Tamara L. Berg 2 Luis E. Ortiz 1 1 Stony Brook University 2 University of North Carolina at Chapel Hill Popularity Tags Image Date Social data 26 Friends 29 Following 552 Followers Comments • kool look, love the hair • Great look! • great ... 10 0 10 1 10 2 10 3 10 4 10 0 10 1 10 2 10 3 10 4 10 5 Popularity measure Number of posts votes comments bookmarks Predicted popular Predicted NOT popular Application • Automatic style feedback • Estimating unbiased rating • Web-traffic estimation • Trend analysis Can we reason about the votes? www.chictopia.com In-network popularity Out-of-network popularity 0 5 10 15 20 25 0 200 400 600 Binary TopïK • Chictopians’ votes • Social network • 328K posts • Crowdsourced votes (AMT) • No social network • 3,000 posts x up to 25 votes Modeling posts – how much does each factor matter? 10 0 10 1 10 2 10 3 10 4 10 5 10 6 10 0 10 1 10 2 10 3 10 4 10 5 Number of connections Number of users friends followers followees Social factors Content factors Other • User identity • Previous posts • Node degrees • Date bias • Tag TF-IDF • Color entropy • Image composition • Style descriptor mean-std pooling key points State-of-the-art computer vision feature RGB color Lab color MR8 texture Boundary distance Skin-hair detection 328K fashion blog posts In-network Out-of-network Regression R 2 Prediction accuracy Reg. R 2 Prediction accuracy Votes Log-votes Top-25% Top-75% Top-25% Top-75% Social 0.372 0.491 0.847 0.779 0.423 0.845 0.787 Content 0.142 0.252 0.782 0.736 0.395 0.869 0.836 Both 0.341 0.493 0.844 0.776 0.475 0.884 0.857 • Significant social influence in network but not necessarily out of network. • Are social and content evaluating different aspect of posts? • Predicting top pictures are consistently easier than not popular pictures.