With the rapid development of online video news, how to accurately predict audience behavior has become an important issue for optimizing content recommendations and improving user experience. This paper introduces the Long Short-Term Memory (LSTM) model to more accurately predict user viewing behavior. First, this paper constructs a dataset of audience behavior characteristics, covering multiple dimensions such as viewing time, click frequency, and dwell time. Then, the LSTM model is used to capture the temporal characteristics of audience behavior, focusing on solving the problem that traditional models are difficult to handle with long-term dependencies. Finally, the LSTM model is used to train and verify the audience's behavior patterns and evaluate its performance in different audience groups. Experimental results show that compared with traditional prediction methods, the model exhibits higher accuracy and stability in capturing the temporal dependency of audience behavior. In the prediction performance evaluation of the LSTM model, in terms of MSE (Mean Square Error), sports news performed best (0.02), followed by entertainment news, and current affairs news performed worst (0.05), indicating that sports news has the highest prediction accuracy. The LSTM model can effectively solve the temporal dependency problem in the prediction of online video news audience behavior, and has strong generalization ability in the context of large-scale data, which is suitable for prediction of various news scenarios.