Leveraging Noisy Lists for Social Feed Ranking
Keywords:social networks, information retrieval, social feed ranking, context collapse, channel collapse
Active users of social networks are subjected to extreme information overload, as they tend to follow hundreds (or even thousands of other users). Aggregated social feeds on sites like Twitter are insufficient, showing superfluous content and not allowing users to separate their topics of interest or place a priority on the content being pushed to them by their “friends.” The major social network platforms have begun to implement various features to help users organize their feeds, but these solutions require significant human effort to function properly. In practice, the burden is so high that most users do not adopt these features. We propose a system that seeks to help users find more relevant content on their feeds, but does not require explicit user input. Our system, BUTTERWORTH, automatically generates a set of “rankers” by identifying sub-communities of the user’s social network and the common content they produce. These rankers are presented using human-readable keywords and allow users to rank their feed by specific topics. We achieve an average top-10 precision of 78%, as compared to a baseline of 45%, for automatically generated topics.