Predicting Pinterest: Organising the World's Images with Human-machine Collaboration

Nishanth Sastry

Abstract: The user generated content revolution has created a glut of multimedia content online -- from Flickr to Facebook, new images are being made available for public consumption everyday. In this talk, we will first explore how, on sites such as Pinterest, users are bringing order to this burgeoning collection by manually curating collections of images in ways that are highly personalised and relevant to their own use. We will then discuss the phenomenon of social bootstrapping, whereby existing mature social networks such as Facebook are helping bootstrap engaged communities of content curators on external sites such as Pinterest. Finally, we will demonstrate how the manual effort involved in curation can be amplified using a unique human-machine collaboration: By treating the curation efforts of a subset of users on Pinterest as a distributed human computation over a low-dimensional approximation of the content corpus, we derive simple yet powerful signals, which, when combined with image-related features drawn from state-of-the-art deep learning techniques, allow us to automatically and accurately populate the personalised curated collections of all other users.