Fake news refers to inaccurate or deceptive information that is portrayed as legitimate news. It is intentionally generated and disseminated to mislead the public. Fake news takes on multiple forms, including altered visuals, invented narratives, and misrepresented accounts of actual occurrences, although this work focuses solely on textual content. Initially, the focus of this work is to evaluate various pre-processing techniques involved in fake news detection, such as TF-IDF, GloVe, and Integer Encoding. Each of these techniques has its own way of converting text to numerical format. Despite numerous studies in this field, there is still a research gap regarding the comparative analysis of TF_IDF (Term Frequency Inverse Document Frequency), Integer Encoding, and GloVe (Global Vector for Word Representation) specifically for fake news tasks. This study aims to bridge this gap by evaluating and comparing the performance of these three popular preprocessing techniques. Next, three RNN variants are used in this experiment for the classification task. They are SimpleRNN (Simple Recurrent Neural Network), LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit). The reason behind choosing RNN variants is RNN is capable of capturing long term dependencies. It is proven to be effective in handling sequential data. It consists of memory that stores the previous important content. GloVe showed high accuracy in GRU model, and it also used only less computational resources, but LSTM took more time and required more computational resources. The results produced by GRU and LSTM for GloVe were better than the rest of the combinations. Integer Encoding also produced good results. But TF-IDF gives poor results when fed to Deep Learning models like RNN, LSTM, and GRU, but when it is fed to Machine Learning Model it gives good accuracy. This is due to sparse matrix generation based on the importance of term frequency. The findings highlight the advantages and limitations of each algorithm, providing valuable guidance for researchers and practitioners in choosing the suitable method for their specific needs. The experimental finding of this work is that GloVe with GRU produces the highest accuracy of 92.15%

GloVe TF-IDF RNN GRU LSTM NLP