Information Diffusion In Online Social Network Based On Deep Representation Learning

With the emergence of online social networks(OSNs),the way people create and share information has changed,which becomes faster and broader than traditional social media.Understanding how information(both good and harmful)spreads through OSNs,as well as what elements drive the success of information diffusion,has significant impli-cations for a wide range of real-world applications.This dissertation analyzes information diffusion based on deep representation learning.Specifically,this dissertation first stud-ies information cascades modeling and prediction,and then studies the rumor detection on OSNs based on diffusion analysis.The main research contents of this dissertation are as follows:(1)A recurrent cascades convolutional network(Cas CN)is proposed,this model utilizes a combination of graph convolutional network and recurrent neural network to predict the incremental size of a cascade by extracting structural-temporal features from a sequence of timestamp-based subcascade graphs.Besides that,introduces cascade lapla-cian into Cas CN,which overcomes the limitations of graph neural networks when dealing with directed graphs.The experimental results conducted on two real-world datasets,show that Cas CN is well-suited for modeling structural-temporal features from cascades,and the performance on cascade incremental size prediction is better than other baselines.(2)A multi-scale cascades model(MUCas)is proposed,which aims at capturing multi-scale features for information cascade incremental size prediction.MUCas uti-lizes a multi-scale graph capsule network and an influence attention to learn and fuse the dynamic-scale,direction-scale,position-scale,and high-order-scale information.In addition,designs a time interval-based subcascade graph sampling method to improve the sampling efficiency and quality.In experiments,MUCas is demonstrated particularly ef-fective at extracting features on cascades from different scales,and multi-scale features are vital for improving prediction accuracy.(3)A macroscopic and microscopic-aware rumor detection model(MMRD)is pro-posed,this model detects rumors by only exploring different levels of diffusion patterns.MMRD leverages graph neural networks and bidirectional recurrent neural networks to learn the macroscopic diffusion pattern and microscopic diffusion pattern from the diffu-sion graph and diffusion path,respectively.MMRD also leverages the knowledge distil-lation technique to create a more powerful student model and further improve the model’s performance.MMRD is evaluated on the Twitter datasets,its performance is better than other baseline methods and could obtain good detection results in the early stage of rumor spreading.(4)From the user perspective,a user-level rumor detection model(UMLARD)is proposed.UMLARD uses three view-specific embedding methods with distinct inputs to capture user multi-view features,which solves the problem of input features entangled with learned high-level features.UMLARD also innovatively proposes a capsule-based attention layer to replace the multi-head attention mechanism in existing methods,which is more effective in both performance and time cost.The extensive experimental evaluations on several real world datasets demonstrate the effectiveness and robustness of UMLARD in comparison to the state-of-the-art baselines.