Hierarchical Transformer-Based Attention Prediction for User Focus Modeling in Digital Media

Abstract

In the digital media environment, user attention is influenced by various factors, such as content features, temporal features, etc. These features have heterologous characteristics. Under the influence of this feature, it is impossible to accurately mine the potential features of user attention in the time domain, resulting in insufficient prediction accuracy. Therefore, this article proposes an attention mechanism prediction algorithm based on Transformer enhanced hierarchical attention network to address the challenge of rapid shift of user attention focus in digital media environment. By constructing a hierarchical attention network that combines content level self attention and temporal level recurrent attention, key behavior nodes and time-dependent features in user interaction sequences are dynamically captured. Based on the Transformer framework, the multi head attention mechanism is optimized to achieve end-to-end training for focus prediction. The experiment was conducted on four real datasets: Frappe, MovieLens, Criteo, and Avazu. The results showed that the algorithm performed well in prediction accuracy related indicators, reaching an AUC value of 0.9848 on the Frappe dataset, an average improvement of 15% -20% compared to the comparison method. It can respond to changes in user interests in real time and provide accurate decision support for personalized content recommendation and platform operation optimization.