Evaluating Important Factors and Effective Models for Twitter Trend Prediction

Trend prediction for social media has become an important problem that can find wide applications in various domains. In this work, we investigate two basic issues in Twitter trend prediction, i.e., what are the important factors and what may be the appropriate models. To address the first issue, we consider different content and context factors by designing features from tweet messages, network topology, and user behavior, etc. To address the second issue, we investigate several prediction models based on combinations of two fundamental model properties, i.e. (non-)linearity and (non-)state-space. Our study is based on a large Twitter dataset with more than 16 M tweets and 660 K users. We report some insightful findings from comparative experiments. In particular, it is found that the most relevant factors are derived from user behavior on information trend propagation and that non-linear state-space models are most effective for trend prediction.