Factorization Meets Memory Network: Learning to Predict Activity Popularity

We address the problem, i.e., early prediction of activity popularity in event-based social networks, aiming at estimating the final popularity of new activities to be published online, which promotes applications such as online advertising recommendation. A key to success for this problem is how to learn effective representations for the three common and important factors, namely, activity organizer (who), location (where), and textual introduction (what), and further model their interactions jointly. Most of existing relevant studies for popularity prediction usually suffer from performing laborious feature engineering and their models separate feature representation and model learning into two different stages, which is sub-optimal from the perspective of optimization. In this paper, we introduce an end-to-end neural network model which combines the merits of Memory netwOrk and factOrization moDels (MOOD), and optimizes them in a unified learning framework. The model first builds a memory network module by proposing organizer and location attentions to measure their related word importance for activity introduction representation. Afterwards, a factorization module is employed to model the interaction of the obtained introduction representation with organizer and location identity representations to generate popularity prediction. Experiments on real datasets demonstrate MOOD indeed outperforms several strong alternatives, and further validate the rational design of MOOD by ablation test.