Predicting the active period of popularity evolution: A case study on Twitter hashtags

Abstract

The active period of popularity evolution indicates how long online content receives continuous attention from people. Although predicting popularity evolution has largely been explored, researches on predicting active period still remain open. If we know the duration of active period ahead of time, caching systems, online advertising, etc. can run more effectively. Therefore, predicting active period is of great importance, but it is a non-trivial task because of the two major challenges. First, numerous factors can influence the duration of active period. To predict active period accurately, it's difficult to consider what factors and how to embed them in DNN model. Second, the triggering time to predict different active periods must be decided carefully, because the durations of active periods differed from one another. This paper addresses these two challenges, focusing on Twitter hashtags as a case study. To deal with the first challenge, a DNN-based prediction framework is proposed, embedding dynamic and static factors by using LSTM and CNN respectively. To deal with the second challenge, an appropriate value of cumulative popularity is set to trigger predicting active period. Experimental and comparative results show the superiority of our prediction solution, comparing with spikeM and SVR.