A Comprehensive Review and Evaluation of Deep Learning Methods in Social Sciences

Absztrakt

Deep learning (DL) has emerged as a cutting-edge data-driven methodology, revolutionizing the field of social sciences by providing profound insights. The application of DL techniques to model social science big data has resulted in significant discoveries and a rapid transformation of traditional methodologies. In this study, we aim to systematically review and evaluate the performance of DL methods in the social sciences. To ensure a rigorous and efficient exploration of relevant databases, we adhere to the PRISMA guidelines. Publications were sourced from Scopus and Web of Science (WoS). The search syntax encompassed essential DL methods, such as convolutional neural network (CNN), Long short-term memory (LSTM), deep neural network (DNN), deep belief network (DBN), Recurrent neural networks (RNN), and deep reinforcement learning (DRL), specifically applied to the social sciences. We utilized a comprehensive search filter to focus on the DL section and its various applications in the social sciences. These applications were categorized into twelve domains, including social information, social network analysis, social development, social movements, social inequalities, social cooperation, social conflict, social technology, social health, social risk, social environment, and social media. To evaluate the performance of DL models, we analyzed the evaluation metrics reported in each study. A normalized reliability score was calculated to facilitate a uniform evaluation of models across different applications. Furthermore, we employed a relief feature selection technique to identify the most influential parameter affecting the reliability score of DL techniques in social science applications. Our findings suggest that evaluation criteria play a crucial role in determining the effectiveness of DL models, while the influence of the database and application type is moderate. However, certain limitations were identified within the studies reviewed. One prominent limitation is the lack of reporting evaluation criteria values during the evaluation phase, which hinders a comprehensive assessment of the employed models...