A Research Of Information Diffusion Models And Popularity Prediction Using Graph Neural Networks

With the rapid development of the Internet and mobile devices,understanding in-network information diffusion in social networks attracts much attention in both academia and industry,becoming a fundamental research problem in many real-world applications of social network analysis.Popularity prediction aims to predict the final diffusion range or size after observing the early-stage evolution of the information cascade.How to accu-rately predict the popularity of the information cascade in complex,rapidly evolved,and vulnerable to be influenced information network becomes one of the most challenging problems in this field.Most of the existing prediction models rely either on hypothesized stochastic processes,artificially designed feature engineering,or end-to-end deep neural networks.These models achieved some extent of success in information cascade pop-ularity prediction,however,facing several notable challenges:(1)only consider the local structures of information cascades,they cannot simultaneously capture the underlying global and local structures;(2)use simple temporal or structural modeling techniques,they cannot jointly model these characteristics in hierarchy;(3)cannot capture the variations and uncertainties in information diffusion;(4)they cannot utilize unlabeled information cascade data;(5)existing data augmentation techniques cannot be implemented in information cascade graphs;(6)they rely on massive labeled data,which generalizes poorly;(7)the learned information cascade representations cannot be transferred across different information cascade datasets and prediction tasks.In order to address challenges(1-3),we propose Cas Flow,which is a probabilistic popularity prediction framework based on graph neural network and variational inference.It conducts non-linear hierarchical analysis on information cascade graphs and models the variations and uncertainties of information diffusion in social networks.Cas Flow allows efficient information diffusion inference and models the diffusion process by learning the latent representations of both temporal and structural characteristics of information cas-cade graphs.Cas Flow is a pattern-agnostic model leveraging the variational auto-encoders and normalizing flows to learn node-level and cascade-level latent influence factors and uncertainties.Cas Flow has a better prediction performance and robustness.In addition,to address the challenges(4-7),we propose CCGL,which is a prediction model based on graph contrastive self-supervised learning.In particular,it first learns general representations of information cascade graphs by a task-agnostic contrastive self-supervised pre-training on both labeled and unlabeled data.Then it fine-tunes the model by utilizing labeled data in a task-specific manner.It finally uses a teacher-student network for knowledge distilling and transfer learning,which effectively address the “negative transfer” issue.CCGL model simulates the diffusion process of information in social net-works and designs a novel data augmentation strategy Aug SIM for information cascade graph.CCGL can effectively alleviate the overfitting problem when training on small datasets,and possesses a better generalization capability.CCGL can learn general knowl-edge from information cascade data,the learned knowledge can be transferred to other information cascade datasets and prediction tasks for performance improvements.The“unsupervised pre-training,fine-tuning,and knowledge distilling” paradigm of CCGL is a promising direction for the design of future information cascade prediction models.At last,this thesis conducts extensive experimental evaluations in several public large-scaled information cascade datasets.Compared to state-of-the-art baselines,both of the proposed prediction models decreased the prediction errors significantly.