Multimodal Context-Aware Recommender for Post Popularity Prediction in Social Media

Millions of multimodal posts are uploaded, shared, viewed and liked every day in different social networks, where users express their opinions about different items such as products and places. While, some user posts become popular, others are ignored. Even different posts related to the same items shared by different users receive different number of likes and views. Existing research on popularity prediction aggregate all user posts related to different items without considering the preferences of individual user for the items in training a popularity model. This often results in limited success. We hypothesize that popularity of posts differs from one user to the other user, one item to the other items, and posts related to the similar users or similar items may be received the same number of likes. In this paper, we present an approach for predicting the popularity of user posts by considering preferences of individual users to the items. We factorize the popularity of posts to the user-item-context and propose a multimodal context-aware recommender for predicting the popularity of posts. Using our proposal we have the ability of predicting the popularity of posts related to different items which are shared by a specific user. Moreover we are able to predict the popularity of posts shared with different users for a specific item. We evaluate our approach on an Instagram user posts dataset with over 600K posts in total related to different touristic places, as items, in The Netherlands for the task of popularity prediction.