Fake News and Imbalanced Data Perspective

Introduction

“Fake news” has gained a significant research attention worldwide since 2017 (Allcott & Gentzkow, 2017). Although fake news is in the public realization, it is important to note that digitally generated text and content can reach far then expected more than the true news. News or media articles online are less authenticated than that of original news media sources such as magazines or newspapers (Xiao, 2018; Zhang et al., 2019). Massive data is getting produced, either manually or by AI, to have a political or financial gains (Vosoughi et al., 2018; Berinsky, 2015). Fake news is those news stories that claim to be accurate, but contain factual misrepresentations with the intention of arousing emotions, attracting viewership, or deceiving (Mihaylov et al., 2018).

By 2022, more fake information than real information would be processed by most people in developed nations also (Gartner, 2018). This will create imbalance in data in online media ecosystem making people exposed to more of Fake News then Real News.

In recent years, there has been significant contribution in the fake news detection area. Major contributions in this field focus on the detection methodology i.e. feature engineering, model construction and so on. However, one area is yet to be explored in the perspective of fake news i.e. nature of the data at disposal for classification. A severe class disparity among these groups causes one of the major problems faced by current machine learning models used for fake news detection. Therefore, most models struggle to distinguish instances that fall into minority groups correctly. The detection of fake news is difficult because of the infinite magnitude and imbalance nature of news data information. The class imbalance problem in data mining is the biggest problem. The problem of imbalance exists where one of the two groups has more samples than other classes. The algorithm mostly focuses more on classifying major samples while ignoring minority samples or misclassifying them. Minority samples are those that occur occasionally, but are very important. There are various ways to deal with imbalance. In this paper we present evaluation of data oriented methods for imbalance handling i.e. over sampling and under sampling. The goal is to recognize the shortcomings of current approaches to class imbalance; more precisely, the techniques of machine learning.

Many contributions have been dedicated to the issue of classification of imbalance data. In an extensive bulk of literature (Zhi-Hua Zhou & Xu-Ying Liu, 2006; Kaur & Gosain, 2018; Wasikowski & Chen, 2010), many solutions have been proposed that are based on machine learning and data mining algorithms. Class imbalance, however, has remained an unresolved problem (Richhariya & K Singh, 2014; Shuo Wang & Xin Yao, 2012).

In this research, comparative analysis of performance of classifier for imbalance data using is made. Focus is more specifically on data oriented methods with various machine learning algorithms. A research on state-of-the-art technologies currently used in applications was conducted to resolve the issue of imbalance classification. The area for imbalance data chosen here for experimentation is Fake News. Contribution of the work in this book chapter are highlighted in the next section.