Credibility of images using textual techniques for Social Media Platforms

Abstract

Finding credible information is of paramount importance in the digital-age where massive flow of information is perceived by every internet user. Information shared on social media got potential to manipulate the thoughts and perception of masses which could result in controlled deviation of humans behaviour at large. Images and videos paired with false textual data are often used to spread false information within social media. Finding false location context within such image descriptions is a difficult task. Much of research work is focused towards finding image-location credibility using associated textual data. However, there is a need of having an social-media-eccentric approach where previously ignored social media meta-information can be utilized for the prediction of image-location credibility score based on description or keywords shared with image. We proposed a holistic view based approach which revolves around the fact that quality of information being shared to a user within social media is directly proportional to the number of users one is in connection with. This led to a proposed method where one can factor-in user credibility parameters such as age, posts, previous credibility score along with the description (keyword) based clustering and matching with similar images to find estimated score. By merging features from multiple approaches, we are able to closely match the prediction rates when compared to conventional location finding methods. For evaluation and verification, hold-out-validation approach is used. In this work, 77%, 66%, 71% accuracy (±5%ofgroundtruth) has been achieved for 11000, 22000 and 43000 images within Div150Cred data-set while raising the F-Score for 5-7% when compared to existing methods.