Enhancing User Understanding with Big Data: A Comparative Study of Deep Learning and Statistical Methods for Forecasting Online Page Views

In the age of big data, the massive online user activity across desktop and mobile platforms generates an immense volume of web traffic logs. Analyzing and forecasting user behaviors, particularly page views, are vital for organizations aiming to enhance personalization, recommendation systems, and search engine optimization efforts. While traditional statistical methods have long been employed for web traffic forecasting, recent advancements in deep learning offer new opportunities for more accurate and insightful predictions. This paper presents a comprehensive comparative study of traditional statistical forecasting techniques and state-of-the-art deep learning methods applied to publicly available Wikipedia web traffic data, which includes hundreds of thousands of pages with over two years of historical page views. We evaluate various deep learning architectures encompassing different model structures, including recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), multilayer perceptrons (MLPs), transformer-based models without pre-training, and pre-trained foundational models. Our findings reveal that deep learning approaches, particularly those leveraging cross-learning and transfer-learning capabilities, significantly outperform conventional methods in forecasting accuracy. These advanced models provide a powerful means to better understand online users’ browsing activities. The enhanced predictive performance of deep learning frameworks equips data scientists and researchers with more effective tools, ultimately improving productivity and efficiency in the analysis of web traffic patterns.