Fake news detection using natural language processing and machine learning: a comparative study of supervised algorithms and text representation techniques

Abstract

The widespread phenomenon of fake news on digital platforms causes serious social issues, highlighting the need for precise and scalable detection methods. There was a comparison among studies being carried out. It is responsible for assessing the functional strategies of supervised machine learning algorithms and text representation. The main purpose of this study was seeking the most effective feature extraction tech niques and algorithmic integration to improve how well classification performed. With the supportive source of comprehensive evaluation methodology and pertinent peer-reviewed publications, databases were utilised to searching for phrases like “supervised learning”, “fake news detection”, “Bag-of-Words”, and “word-embeddings”. Some highlighted examples are IEEE Xplore, ScienceDirect, and SpringerLink. Bag-of-Words (BoW) and word-embeddings (Word2Vec, GloVe) were approached as the two main subjects of the investigation. By contrast, combining them with supervised algorithms like Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and Naïve Bayes (NB) had a conflict. The arrangement of comparative tables was deployed for datasets, feature extraction techniques, classifiers, and perfor mance indicators to generate codes. Through the result, it was recorded that the embedding-based tech niques had a greater performance than BoW. The more sophisticated combination of classification systems like SVM and collaborative models were, the more efficient majority of scenarios was. Previous trained em beddings were required with above standard lexical models regarding to accuracy and F1-score through data collections like LIAR and FEVER. It was undeniable to realise how majorly the degree of preprocessing, algorithmic sophistication, and qual ity of data collection impacted on identifying performance. Under any encouragement of results, some drawbacks still were existent, including inconsistent assessment criteria, and unbalanced dataset, and dif ferent preprocessing techniques used in different research. The best condition for developers to choose the suitable models depended on how available data is and computing capacity is one example of the practical ramifications, indeed. The suggestion of future studies brought the new investigation of deep learning and unsupervised learning techniques and create uniform data points for more reliable assessment.