Article Preview
Top1. Introduction
Recently, online social networks (OSNs) are becoming a convenient and popular source of latest news since OSNs make peer-to-peer communication easier, open, and instant. Such an open and unmoderated discussion has led OSNs a fertile land for unverified information spread on a larger scale. This unverified information is otherwise called as unconfirmed information or rumor. Rumors permeated at every corner of social networking application due to large-scale dissemination in a brief time. This kind of huge circulation of rumors leads to potential damages and elevated societal harms across online and offline social communities. Detecting rumors and controlling at the earliest is imperative to minimize the damages to the network. Consequently, the accurate identification of rumors in OSNs is highly beneficial and desirable.
In the social psychology field, the rumor is defined as a piece of information whose source cannot be verified as true or false at the time of circulation (Allport & Postman, 1947). In OSN terms, a rumor is a story or claim that is unverified and being propagated among participants in the network. Research task in this area of study is to identify and control the rumors using various features of every information. The prominent features that determine the veracity of rumors are the temporal, the linguistic, and the structural aspects of the information (Kwon, Cha, Jung, Chen, & Wang, Prominent features of rumor propagation in online social media, 2013). Currently, a rich set of principled rumor detection methods are proposed for online social networks which use content, temporal and structural features of the information (Castillo, Mendoza, & Poblete, Information credibility on twitter, 2011) (Kwon, Cha, & Jung, Rumor detection over varying time windows, 2017) (Liu, Nourbakhsh, Li, Fang, & Shah, 2015). Few types of research leverage the wisdom of users (Liu, Nourbakhsh, Li, Fang, & Shah, 2015) (Zhao, Resnick, & Mei, 2015), propagation network (Ma, Gao, & Wong, Rumor Detection on Twitter with Tree-structured Recursive Neural Networks, 2018) and so on. However, the fundamental problems such as input uncertainty, input noise, data sparsity, and, output interpretation of this dynamic and complicated information propagating network, OSNs, are not well addressed by these state-of-the-art methods.
OSNs are driven by information sharing across the participants in an uncontrolled environment. Participants can interpret and share the same information in different semantical directions which introduces uncertainty and inconsistency in information in its further spreads in the network. Such imprecise information needs to be handled while identifying rumors to avoid misleading classification results. Apart from this, when rumor needs to be identified earlier in its propagation, the common problem would be sparsity of input which in-turn contains more noise in data (Haufe et al., 2014) (Ma et al., 2016). Such a very less, noise incorporated, data often does not help in rumor identification. This situation causes the earlier detection job a failure. Also, machine-learning and deep-learning classifications are commonly having the problem of output interpretation (Lipton, 2016) (Lou, Caruana, & Gehrke, 2012). The output generated by these models does not have a proper interpretation as to how the results arrive. This leads to the problem of not knowing how the presented model works.