Affective Signals in a Social Media Recommender System

Abstract

People come to social media to satisfy a variety of needs, such as being\ninformed, entertained and inspired, or connected to their friends and\ncommunity. Hence, to design a ranking function that gives useful and\npersonalized post recommendations, it would be helpful to be able to predict\nthe affective response a user may have to a post (e.g., entertained, informed,\nangered). This paper describes the challenges and solutions we developed to\napply Affective Computing to social media recommendation systems.\n We address several types of challenges. First, we devise a taxonomy of\naffects that was small (for practical purposes) yet covers the important\nnuances needed for the application. Second, to collect training data for our\nmodels, we balance between signals that are already available to us (namely,\ndifferent types of user engagement) and data we collected through a carefully\ncrafted human annotation effort on 800k posts. We demonstrate that affective\nresponse information learned from this dataset improves a module in the\nrecommendation system by more than 8%. Online experimentation also demonstrates\nstatistically significant decreases in surfaced violating content and increases\nin surfaced content that users find valuable.\n