This paper proposes a deep learning model based on multi-layer LSTM. By constructing a multi-dimensional input feature system, covering event attributes, user interaction and temporal context information, it realizes dynamic modeling of viewers’ behaviors such as entering, staying and interacting. Based on the data of 1,200 actual events, the experiment is designed to compare the traditional model with the deep model and analyze the feature ablation. The results show that the proposed model outperforms the comparison method in Accuracy, F1-Score and AUC, and exhibits stronger generalization ability and practicality.