Information Diffusion Prediction Based on Deep Attention in Heterogeneous Networks

Abstract

Understanding how information is spread is critical necessities in many real-world application domains, so the study of information diffusion in social networks has attracted considerable research interest. Compared with the widely studied homogeneous networks, heterogeneous networks can more accurately model the process of information diffusion with multiple channels and more closely match the pattern of information diffusion in the real world. However, the complex structural information and rich semantic information in heterogeneous networks bring challenges to the extraction and utilization of effective information. In addition, the existing heterogeneous diffusion models do not fully consider the different effects of different diffusion channels on information propagation. Therefore, we propose a Heterogeneous Deep Attention Diffusion model (HDAD). HDAD first extracts the information matrix that integrates the network topology and information diffusion state based on the meta-path for simplifying the network and retains the effective information of the network; Secondly we design a deep learning architecture to learn low-dimensional embeddings and capture non-linear relationships; and then attention mechanism is used to learn the importance of different diffusion channels and to combine the low-dimensional embeddings under different semantics in the network rationally. Experiments on the public DBLP and ACM datasets are conducted, and the experimental results show that HDAD can fully exploit the information in the network and the prediction performance is better than the existing models.