A Comparative Study of Machine Learning Algorithms for the Detection of Fake News on the Internet

Abstract

Context: The increase in the proliferation of fake news on the Internet has significantly impacted the quality and veracity of information received by society. Problem: The malicious use of information can compromise democracy by manipulating people’s opinions. In addition, there are few facilitating mechanisms that classify and help the citizen to know whether a certain news propagated is true or not. This problem has driven new research directions in an attempt to classify and identify these news. Methodology: This work in its methodology performs a comparison of algorithms to serve as an intelligent solution in the detection of fake news in Portuguese. About 12,000 news featured the dataset used for this analysis. Pre-processing techniques were used to analyze the patterns of these news, as well as to reduce noise and eliminate null information. The algorithms used for comparison were Logistic Regression, Stochastic Gradient Descent, Support Vector Machine and Multilayer Perceptron. Result: The results obtained showed that the models generated by the four algorithms obtained an accuracy greater than 90%. To choose the best algorithm, metrics such as precision, recall and f-measure were used for each of the models. The SVM algorithm had the best performance, with 96.39% accuracy. Contribution: In addition to the analytical results presented, this work brought as contributions the availability of a database containing news in Portuguese and an analysis, from the text of the news, both grammatical and structural, in order to detect the existing patterns between true and false.