Deep Learning Methods on Predicting Social Hotspots with Attention Mechanism: A Case Study on Zhihu

Topic hotspot prediction technology not only allows researchers to know future research hotspots, but allows social organizations to know the issues of public concern. Although deep learning has achieved impressive results in hot spot prediction, there are still some challenges such as the topics popularity of public opinion was greatly affected by external news, it often fluctuates violently at particular times, resulting in insufficiently accurate predictions. To timely and accurately analyze the attention and demands of online topics, an Att-BLSTM model is optimized to cluster the popularity and predict its change trend based on millions of hot topics and their reply information. First, a K-means algorithm is used to cluster the hot topics of the hot topics. with the aim of more accurately describe the detailed features and long-term trends of topic popularity, a prediction model based on attention mechanism Bidirectional Long Short-Term Memory (BiLSTM) network is optimized, called as Att-BLSTM model. A large number of numerical experiments were conducted using information on hot topics from the Chinese Zhihu website. Depending on the results of the evaluation metrics, the comparative excellence of the Att-BLSTM model is demonstrated.