Chaotic Neural Architecture Search with Cosine Similarity for Accurate Fake News Detection

Abstract

Online social network platforms are more accessible than traditional news sources, and published content can reach large audiences very quickly. These types of news, which are often shared without checking whether they are real or not, can cause reputation assassination to individuals, as well as chaos, economic losses, fears and concerns that can affect societies. Therefore, early detection of fake news is an important step in preventing such situations. In this study, we propose a deep model based on neural architecture search for fake news detection. The method used in the study has two important original features. The first is that it uses the chaotic random architecture search technique instead of the uniform random architecture search technique. This allows flexible jumps in the search space, especially in the exploration phase. The second feature is that it prefers a population management with dynamically sized and diversity mechanism so that the search space can be scanned homogeneously. The diversity mechanism is based on cosine similarity and helps to make the hyperparameters of the candidate architectures as different as possible. The proposed method, based on the CNN model, uses a non-deterministic metaheuristic approach for the random neural architecture search. Comparative tests of the model have been performed on the coronavirus and Syrian war data sets. The performance of the proposed method was benchmarked against a comprehensive suite of nine models, including classical machine learning algorithms, standard deep learning architectures, and state-of-the-art Transformer models such as BERT and RoBERTa. The results demonstrate the success and efficiency of the proposed method, particularly in data-scarce scenarios.