Predicting Popular and Viral Image Cascades in Pinterest

Abstract

The word-of-mouth diffusion has been regarded as an important mechanism to advertise a new idea, image, technology, or product in online social networks (OSNs). This paper studies the prediction of popular and viral image diffusion in Pinterest. We first characterize an image cascade from two perspectives: (i) volume - how large the cascade is, i.e., total number of users reached, and (ii) structural virality - how many users in the cascade are responsible for attracting other users. Our model predicts whether an image will be (a) popular in terms of the volume of its cascade, or (b) viral in terms of the structural virality. Our analysis reveals that a popular image is not necessarily viral, and vice versa. This motivates us to investigate whether there are distinctive features for accurately predicting popular or viral image cascades. To predict the popular or viral image cascades, we consider the following feature sets: (i) deep image features, (ii) image meta and poster's information, and (iii) initial propagation pattern. We find that using deep image features alone is not as effective in predicting popular or viral image cascades. We show that image meta and poster's information are strong predictors for predicting popular image cascades while image meta and initial propagation patterns are useful to predict viral image cascades. We believe our exploration can give an important insight for content providers, OSN operators, and marketers in predicting popular or viral image diffusion. © Copyright 2017, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.