Popularity prediction of online contents is always a tool of emergency management, business decision-making, and public opinion monitoring. Most previous work has made efforts to predict the volumes or levels of popularity, but patterns of popularity evolution are remaining largely unexplored. Actually, topic popularity patterns can offer more detailed information for event detection and early warning. In this article, we proposed an effective method to discover and predict the popularity patterns of topics on the Internet which combined clustering and classification models. This method does not rely on the early time data of topic propagation, so it can predict the future popularity pattern at the initial stage of topic releasing. First, we chose a time series clustering algorithm K-SC to obtain basic types of topic popularity patterns. Then, through acquiring and evaluating multiple features related to the topics including publisher features, outward characteristics of content and textual ones, we built the prediction model of topic popularity patterns based on machine learning methods. The experimental results show that it is suitable to cluster four basic patterns of topic popularity from the experimental data. What’s more, making use of certain initial characteristics, Decision Tree model can effectively predict the popularity pattern of a newly released topic, with an accuracy of 89.4%.