ConCas: Cascade Popularity Prediction Based on Topic-Aware Graph Contrastive Learning

Abstract

AbstractCascade popularity prediction on social networks has attracted much attention from scholars over the past few years. Many existing methods take cascade size during the observation period as a key feature. However, when making early predictions where the observation period is short, the difference in cascade size between popular and unpopular contents becomes insignificant. Therefore, this paper proposes a topic-aware graph pre-training model based on the self-supervised learning termed ConCas, which fully utilizes graph structural and content topic information for accurate cascade popularity prediction. Specifically, graph contrastive learning is used on cascade graphs to identify the key structural characteristics that discriminate the popular and unpopular cascades. To further improve the model performance, the underlying social network is used to obtain embeddings as node features. Besides, Latent Dirichlet Allocation (LDA) topic modeling is performed on the message content to generate topic tags, which are further embedded in the cascade graph as node features. Finally, the embeddings of the learned cascade graph are fed to downstream learning models for popularity prediction. Experimental results on the Weibo dataset show that ConCas achieves higher prediction accuracy than the state-of-the-art baselines.KeywordsPopularity predictionGraph neural networksContrastive learningLDA topic modeling