CEAM - The Effectiveness of Cyclic and Ephemeral Attention Models of User Behavior on Social Platforms.

To improve the user experience as well as business outcomes, social platforms aim to predict user behavior. To this end, recurrent models are often used to predict a user's next behavior based on their most recent behavior. However, people have habits and routines, making it plausible to predict their behavior from more than just their most recent activity. Our work focuses on the interplay between ephemeral and cyclical components of user behaviors. By utilizing user activity data from social platform Snapchat, we uncover cyclic and ephemeral usage patterns on a per user level. Based on our findings, we imbued recurrent models with awareness: we augment an RNN with a cyclic module to complement traditional RNNs that model ephemeral behaviors and allow a flexible weighting of the two for the prediction task. We conducted extensive experiments to evaluate our model's performance on four user behavior prediction tasks on the Snapchat platform. We achieve improved results on each task compared against existing methods, using this simple, but important insight in user behavior: Both cyclical and ephemeral components matter. We show that in some situations and for some people, ephemeral components may be more helpful for predicting behavior, while for others and in other situations, cyclical components may carry more weight.