ANALYSIS OF INFORMATION POLLUTION ON WEB AND SOCIAL MEDIA

Abstract

In the era of information overload, restiveness, uncertainty and implausible content all around; information credibility or web credibility refers to the trustworthiness, reliability, fairness and accuracy of the information. Information credibility is the extent up to which a person believes in the content provided on the internet. Every second of time passes by millions of people interacting on social media, creating vast volumes of data, which has many unseen patterns and trends inside. The data disseminating on the web, social media and discussion forums have become a massive topic of interest for analytics as well as critics as it reflects social behaviour, choices, perceptions and mindset of people. A con siderable amount of unverified and unauthenticated information travels through these net works, misleading a large population. Thus, to increase the trustworthiness of online social networks and mitigate the devastating effects of information pollution; timely detection and containment of false content circulating on the web are highly required. To analyse and address the issue of information pollution on web and social media, we have initially reviewed the most popular and prominent state-of-the-art solutions, com pared them and presented. Based on the literature survey, these solutions are categorized and analysed. The prevalent approaches in each modality are studied and highlighted in detail, which helped identify the research gaps in this area. To overcome the issue, our proposed solutions are focused on two categories: semi-supervised textual fake news classification frameworks and supervised multimodal veracity analysis frameworks. The first model developed for semi-supervised textual fake news classification frameworks proposes an innovative Convolutional Neural Network built on the self-en sembling concept to take leverage of the linguistic and stylometric information of anno tated news articles, at the same time explore the hidden patterns in unlabelled data as well. Next, we aim to design a semi-supervised fake news detection technique based on GCN (Graph Convolutional Networks). The recommended architecture comprises of three basic components: collecting word embeddings from the news articles in datasets utilising vi GloVe, building similarity graph using Word Mover’s Distance (WMD) and finally apply ing Graph Convolution Network (GCN) for binary classification of news articles in semi supervised paradigm. In the category of supervised multimodal veracity analysis frameworks, the first model consists of four independent parallel streams capable enough to detect specific for gery formats. All four streams are applied to each input instance. Hierarchical Attention Network deals with headline and body part; Image captioning and headline matching mod ule require all the three parts headline, body and image. Noise Variance Inconsistency and Error Level Analysis focuses only on images accompanied with news text. These inde pendent predictions are finally combined using the max voting ensemble method. The sec ond model aims Inception-ResNet-v2 to extract visual features. The models BERT and ALBERT have been used to elicit textual attributes. Diverse text input forms, like English articles, Chinese articles and Tweets, have been used to make our model robust and usable across multiple platforms. The architecture of Multimodal Early Fusion and Late Fusion has also been experimented with and analysed in detail by applying it on different datasets. Finally, this thesis work is concluded with significant findings and future research aspects in veracity analysis of web and social media information.