Microscopic diffusion prediction based on multifeature fusion and deep learning

Deep learning methods have been widely employed to enhance microscopic diffusion prediction in social networks. However, the existing methods have the problem of insufficient extraction of features in the information dissemination process. For example, these methods do not consider the impact of the propagation chain of the most recently infected nodes on the subsequent propagation of the message or the impact of changes in the neighboring nodes on the propagation path of the message. Therefore, the prediction accuracy is not high. To solve the above problems, describe the information diffusion process from multiple perspectives, and discover more hidden features, this paper proposes a microscopic diffusion prediction framework — multifeature fusion and deep learning for prediction (MFFDLP).

The microscopic diffusion prediction framework is divided into three main parts: extracting the static features from the network topology and the information diffusion sequence, capturing dynamic diffusion characteristics from the information diffusion graph, and predicting the next infected node. (1) First, node embedding and node structure context are extracted from historical friendship graphs and information diffusion sequences. The gate recurrent unit (GRU) is applied to mine the deep global temporal features from the connected vectors. To further enhance the role of the recently infected node, GRU is used to mine the local temporal features from the structure context of the node. These two features are fused to form the information diffusion sequence features. (2) Capture dynamic diffusion characteristics from the information diffusion graph. These features represent changes in users' interaction or interest. An information diffusion graph is built based on the historical information diffusion sequence. The diffusion graph is then divided into subgraphs in chronological order. A graph attention network is applied to capture the node features from each subgraph, and the edge features are aggregated from the node features. Using an embedding lookup method and fusing the nodes and their edge features, the dynamic diffusion characteristics of the users in an information diffusion sequence are obtained. (3) Predict the next infected node. To further analyze the context interaction within the diffusion sequences, a dual multihead self-attention mechanism is applied to separately capture the contextual information from information diffusion sequence features and node dynamic diffusion characteristics. Then, a fully connected layer and Softmax are used to predict the next infected node. Finally, experiments on three real networks show that the proposed method outperforms the state-of-the-art models. The experimental results demonstrate the unique advantages of MFFDLP for microscopic diffusion prediction.

Comparative experimental results on three public datasets show that the proposed method outperforms the comparative methods by up to 9.98% in the accuracy of microscopic diffusion prediction.

This method comprehensively combines the friendship graph, information diffusion sequence, and diffusion graph. Multiple deep learning models are used to extract multiple features from static and dynamic perspectives. Comparative experiments on multiple datasets demonstrate that MFFDLP can mine and fuse multiple features more effectively, thus improving the prediction accuracy of information diffusion.