Temporal Graph Attention Network for Spatio-Temporal Feature Extraction in Research Topic Trend Prediction

Abstract

Predicting trends across multiple research topics presents a significant challenge for multivariate time series forecasting. In the temporal dimension, the fluctuation in popularity of individual research topics reveals trend features, while in the spatial dimension, the interdependence among frequently co-occurring research topics reveals correlation features. However, existing methods often overlook this spatial correlation feature or regard it as static and unchanging, leading to suboptimal prediction results. To address these shortcomings, this paper proposes a novel spatio-temporal graph neural network model, the Temporal Graph Attention Network (T-GAT). This model employs temporal convolution to extract the temporal trend features of research topics and utilizes multi-head graph attention to extract the spatial correlation features among these topics. Notably, the graph attention mechanism is capable of dynamically adjusting the weights of neighboring nodes based on the attention scores learned from the data, thereby reducing the risk of over-smoothing. Experiments conducted on the dataset from the Wanfang Academic Journal Database demonstrate that the proposed model outperforms baseline methods.