Time series modeling and anomaly removal for Wikipedia web traffic prediction

Abstract Accurate forecasting of web traffic is essential for a web service provider who wants to manage infrastructure resources and decision-making processes. Internet users rely heavily on Wikipedia articles for information, and over the past few years, the usage of Wikipedia has increased significantly. Several forecasting and prediction models for specific web page traffic have been published previously that include statistical and deep learning methods. The web traffic data modeling of Wikipedia pages is challenging due to high dimensionality, seasonality, anomalies, and missing values. These anomalies, if left undetected, can adversely affect the overall performance of time series models. In this research, a comprehensive methodology for web traffic prediction is proposed that integrates robust preprocessing, particularly anomaly detection and removal using Isolation Forest, with a comparative analysis of eight time series forecasting models, including ARMA, ARIMA, Auto ARIMA, Exponential Smoothing, Prophet, LSTM, BiLSTM, and MA. The dataset for this research is a Kaggle competition dataset, and Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Error (MAE) are utilized as the evaluation metrics to measure the model’s performance. The findings indicate that the Moving Average model performs better for this dataset when evaluated using RMSE, MAPE, and MAE value (percentage deviation) as compared to other models. A comparison of eight time series model performances is presented based on the Wikipedia dataset, which comprises 145,000 articles.