Abstract:
The technological advances and the expansion in the digital scenario of the world in precious
decade have brought about the capability of rapid information transfer over vast geographical
regions. The rapid transfer of information on one hand has brought relief to human race in
many aspects. At the same time the world as global village face many challenges too due to
this rapid transfer of information. Among the challenges that have accompanied the
revolution in information transfer is the exchange of misinformation. Rapid spreading of
misinformation is a growing concern worldwide as it has the capacity to greatly influence not
only individual reputation but societal behaviours. The consequences of unchecked spreading
of misinformation can not only vary from political in nature to financial, but can also effect
global opinion for a long time. The damage of the menace of fake news, being spread around
in any form, can be beyond the imagination of human mind and its outcomes everlasting for
the generations to come. Thus, detecting fake news is extremely important as well as
challenging because the ability to accurately categorize certain information as true or fake is
limited even in human. Moreover, fake news are a blend of correct news and false
information making accurate classification even more confusing. In this paper, we propose a
novel method of multilevel multiclass fake news detection based on relabelling of the dataset
and learning iteratively. We tested our algorithm on metadata, text and a combination of both.
The proposed method outperforms the benchmark with an accuracy of 39.7% but
maximum accuracy is achieved by holdout method using SVM classifier that is 66%.
Our experiments indicate that profile of the source of information contributes the most in fake
news detection.
