| dc.description.abstract | This thesis aims to improve the accuracy of fake news detection by using Tsetlin Machines (TM). TMs are well suited for noisy and complex relations within the provided data, which on initial analysis, overlaps nicely with characteristics found in fake news. We provide a performant and deterministic preprocessor, which is responsible for tokenizing, lemmanzing, and encoding to a representation that the TM understands. We compare our approach with TMs against Neural Networks (NN) models over a variety of well-known datasets within the fake news domain. Our findings show from comparable results to significant improvements over state of the art. Additionally, we show how TMs allow for interpretable propositional logic rules. For datasets with 2 classifications, we further convey these rules during inference by applying a color between red and green, which shows the intensity and what direction each word pulls the classification towards. | |
