Pinterest Analysis and Recommendations

Pinterest is a visual discovery tool for collecting and organizing content on the Web with over 70 million users. Users “pin” images, videos, articles, products, and other objects they find on the Web, and organize them into boards by topic. Other users can repin these and also follow other users or boards. Each user organizes things differently, and this produces a vast amount of human-curated content. For example, someone looking to decorate their home might pin many images of furniture that fits their taste. These curated collections produce a large number of associations between pins, and we investigate how to leverage these associations to surface personalized content to users. Little work has been done on the Pinterest network before due to lack of availability of data. We first performed an analysis on a representative sample of the Pinterest network. After analyzing the network, we created recommendation systems, suggesting pins that users would be likely to repin or like based on their previous interactions on Pinterest. We created recommendation systems using four approaches: a baseline recommendation system using the power law distribution of the images; a content-based filtering algorithm; and two collaborative filtering algorithms, one based on one-mode projection of a bipartite graph, and the second using a label propagation approach.