Extracting Image Context from Pinterest for Image Recommendation

Image search and recommendation engines try to extract relevant images for a user's information need. Existing approaches use manual tags of networks like Flickr or the surrounding webpages to create context to foster the search. Pinterest as a new upcoming social bookmarking service allows us to gain more context for an image than before. By using board headline, pin descriptions, and the actual content of the bookmarked pages we build a much more complex context. As a use case, we recommend images for blog articles to show the feasibility of the context of Pinterest. We apply tag-based retrieval models to actual propose matching images for article texts. This enables blog authors to get image suggestions for their articles to speed up the creation of appealing articles. Our evaluation shows that a retrieval model based on cosine similarity yields promising results. Given the bookmarked pages, it reaches a precision of 96% to predict the pinned images. Further, a user survey yields that the recommended images are actual usable for the articles.