Search

Online disinformation has become one of the most severe concerns in today's world. Recognizing disinformation timely and effectively is very hard, because the propagation process of disinformation is dynamic and complicated. The existing newest research leverage uniform time intervals to study the multi-stage propagation features of disinformation. However, uniform time intervals are unrealistic in the real world, cause the process of information propagation is not regular. In light of these facts, we propose a novel and effective framework Multi-stageDynamicDisinformationDetection with Graph Entropy Guidance(MsDD) to better analyze multi-stage propagation patterns. Instead of traditional snapshots, we analyze the dynamic propagation network via graph entropy, which can work effectively in finding the dynamic and variable-length stages. In this way, we can explicitly learn the changing pattern of propagation stages and support timely detection even at the early stages. Based on this effective multi-stage analysis framework, we further propose a novel dynamic analysis model to model both the structural and sequential evolving features. Extensive experiments on two real-world datasets prove the superiority of our model. We open the datasets and source code at https://github.com/researchxr/MsDD. [ABSTRACT FROM AUTHOR]

In recent years, online social networks (OSNs) have thoroughly penetrated people's lives. Since information always flows along with various online relations in OSNs, analysing these relations becomes one of the most fundamental problems in various applications. Unfortunately, limited by the privacy concerns, data availability and the pow-law distributions of most OSNs, we can not observe enough relation links all the time, which is difficult to improve downstream tasks. To address this problem, many studies try to predict potential relations in online social networks via existing pair-wise links. However, when the observable pair-wise links are extremely sparse, most of them fail to learn the smoothness on the networks and make their proposed methods brittle. In light of this, we go beyond pair-wise relations and leverage hypergraphs to learn higher relations in the graphs. A hypergraph allows one hyperedge to connect multiple nodes, which is perfect to include more potential pair-wise links and can guarantee smooth node embeddings for better link prediction performance. In this paper, we aim at predicting potential links in sparsely observed networks. To achieve this goal, we first start from some blurry hyperedges and then proposed a novel hyperedge shrinking method to make the learned hyperedges more hierarchical. This method can learn hypergraph structure automatically from the given sparsely observed links and rely less on manual design. Following this, we further propose a novel hypergraph-based graph neural network to learn potential links in the graph. To address semantic fusion in the heterogeneous networks, we put forward multi-level reconstruction methods to preserve both specific semantics denoted by meta-paths, and high-level semantics denoted by hypergraphs. We compare our method with four state-of-the-art baselines. Extensive evaluations demonstrate that our method can achieve the best linking prediction results, especially when the networks are sparse. [ABSTRACT FROM AUTHOR]

Personality analysis has been widely used in various social services such as mental healthcare, recommendation systems and so on because its natural explainability for AI applications in Web intelligence. With the penetration of Web2.0, traditional social researches have gradually turned to online social networks. However, for a long time, personality detection from online social texts has sunk into an embarrassing situation for the lack of large labeled datasets. Limited by supervised learning frameworks and small labeled datasets, prior works mainly detect one's personality in the individual perspective, which may not well meet the challenges of massive un-labeled data in the near future. In this paper, we present a first look into group-level personality detection and we use an unsupervised feature learning method instead of supervised methods used in most related works. We propose AdaWalk, a new and novel model of group-level personality detection by learning the influence from text generated networks. The model uses different kernels to evaluate how much a given node should decide its walk path locally or globally. The advantage of AdaWalk is three-folded: a) the model is an unsupervised feature learning method, which means it relies less on annotations. b) by traversing the network, we can capture the influence in the group level, thus the analysis of one's personality is not only based on individual records but also the information in groups. Therefore, AdaWalk can leverage small datasets more comprehensively. c) AdaWalk is scalable and can be easily transformed as distributed algorithms, which means it has more potential, compared with existing personality detection methods, to meet the massive data without annotations. We use AdaWalk to predict users' Big Five personality scores in FIVE heterogeneous personality datasets. Compared with more than TEN famous related methods, AdaWalk outperforms the others, meanwhile verifying the significance of the group perspective and unsupervised feature learning methods in the application of personality analysis. To make our experiment repeatable, AdaWalk and related datasets are available at https://xiangguosun.strikingly.com. [ABSTRACT FROM AUTHOR]

The rise of online crowdsourcing services has prompted an evolution from traditional spamming accounts, which spread unwanted advertisements and fraudulent content, into novel spammers that resemble those of normal users. Prior research has mainly focused on machine accounts and spams separately, but characteristics of new types of spammers and spamming make it difficult for traditional methods to perform well. In this paper, we integrate the study of these new types of spammers with the study of crowdturfing microblogs, investigating the mechanism of crowdsourcing and the close relationship between crowdturfing spammers and microblogs in order to detect new types of spammers and spams more precisely. We propose a novel semi-supervised learning framework for co-detecting crowdturfing microblogs and spammers by comprehensively modeling user behavior, message content, and users' following and retweeting networks. In order to meet the challenge of sparsely labeled datasets, we design an elaborate co-detection target optimal function to minimize empirical error and to permit the dissemination of sparse labels to unlabeled samples. The advantage of this framework is threefold. First, through a deep-level mining of new-type spammers, we aggregate a number of new-found features that can help us make significant distinctions between normal users and new-type spammers. Secondly, by modeling both following networks and retweeting networks, we characterize the essence of the crowdsourcing mechanism abused by spammers in crowdturfing microblog diffusion to markedly increase detection performance. Thirdly, through our optimal function based on semi-supervised methods, we overcome the problem of label sparseness, thus obtaining a more reliable capacity to deal with the challenges of big, sparsely labeled data. Extensive experiments on real datasets demonstrate that our method outperforms four baselines in various metrics (Precision-Recall, AUC values, Precision@K and so on). We also develop a robust system, the functions of which include data collection and availability analysis, spam and spammer detection, and visualization. To render our experiments replicable, we have made our dataset and codes openly available at https://github.com/sunxiangguo/Crowdturfing. [ABSTRACT FROM AUTHOR]