Modeling Social Media Content with Word Vectors for Recommendation

Ying Ding, Jing Jiang

Abstract: In social media, recommender systems are becoming more and more important. Different techniques have been designed for recommendations under various scenarios, but many of them do not use user-generated content, which potentially reflects users’ opinions and interests. Although a few studies have tried to combine user-generated content with rating or adoption data, they mostly reply on lexical similarity to calculate textual similarity. However, in social media, a diverse range of words is used. This renders the traditional ways of calculating textual similarity ineffective. In this work, we apply vector representation of words to measure the semantic similarity between text. We design a model that seamlessly integrates word vectors into a joint model of user feedback and text content. Extensive experiments on datasets from various domains prove that our model is effective in both recommendation and topic discovery in social media.