Automatic on-device filtering of social networking feeds

Abstract

Many people follow social networking services and find it difficult to locate essential content on mobile devices. Automatic filtering of the feeds is one solution to this problem. A system learns a model for each user, based on metadata (e.g., content types and contacts) and click histories for old feed items, predicts the click probability for incoming items, and automatically filters out less important ones. In this study, we implemented several alternative automatic filtering systems and evaluate their offline accuracy and user acceptance. 40 users completed the evaluation in a field study. Two main findings emerge from the study. Firstly, PageRank and Bayesian predictors are valid methods; an ensemble predictor combining the two further improves the prediction accuracy. Secondly, people show high acceptance of the automatic filtering function. The participants using the filtering function found it easier to access interesting content than did the participants without the filtering. On average, they also felt greater sense of control, due to the reduced feed volume.