Modeling Content Interaction in Information Diffusion with Pre-trained Sentence Embedding

Abstract

Social networks have become indispensable parts of our daily life, and therefore understanding the process of information diffusion over social networks is a meaningful research topic. Usually, multiple pieces of information do not spread in isolation; rather, they interact with each other throughout the diffusion process. This paper aims to quantify these interactions by modeling users' forwarding behavior after reading a series of information. Inspired by several successful components prevalent in recent research of deep learning, i.e., long short term memory (LSTM) network and bi-directional encoder representation from transformers (BERT), we designed IMM Enhanced model and InfoLSTM model. In our experiments on real-world Weibo dataset, both models significantly outperform baselines such as the prior IMM model and IP model, with IMM Enhanced model improving 23.52% and InfoLSTM model improving 32.56% in F1 score (absolute value) compared to that of baseline IMM model. In addition, we visualize the dataset and the parameters learned in IMM Enhanced model, which further enables us to discuss the relationship between text similarity and information diffusion interaction with case studies.