Web Traffic Analysis and Forecasting using Deep Learning Time-Series Approach In case of Commercial Bank of Ethiopia

Abstract

Web traffic forecasting holds immense importance in making data-driven decisions across diverse domains. However, existing studies often rely on Wikipedia datasets that might not fully capture the distinctive aspects of web traffic. Additionally, there is a tendency to prioritize conventional models, overlooking the exploration of potentially superior models. The lack of comprehensive comparisons among different deep learning models hampers our understanding of their relative performance and how datasets impact their effectiveness. These limitations significantly hinder the generalizability of findings, particularly in developing countries. To address these gaps, this paper aims to investigate and compare the effectiveness of six recurrent neural network models by utilizing a unique dataset from local organizations. The objective is to develop a precise web traffic forecasting models by leveraging deep learning techniques and local data, ultimately enhancing decision-making processes. The research process involves stages such as data collection, preprocessing, hyperparameter tuning, model training, prediction, and evaluation. The research paper provides a comprehensive analysis of experiments conducted on web traffic datasets from the Commercial Bank of Ethiopia (CBE) website. The dataset includes visitor counts of web pages, spanning seven years from January 1, 2016, to January 3, 2023, totaling 2560 days of data. To facilitate the analysis, the dataset is divided into training, validation and testing sets. In this study, deep learning techniques, including Long Short Term Memory (LSTM), bidirectional LSTM, bidirectional LSTM with attention, Gated Recurrent Unit (GRU), bidirectional GRU, and bidirectional GRU with attention, were effectively employed to analyze and predict web traffic patterns. The bidirectional GRU with attention model showed great promise, delivering the most impressive results with the lowest Mean Absolute Error (MAE) of 0.06102, Mean Squared Error (MSE) of 0.00713, and Root Mean squared Error (RMSE) of 0.08266. The findings contribute to the understanding of web traffic analysis and display the potential of deep learning in this field. Future work can focus on enhancing and expanding this approach through dataset preparation, model architecture exploration, and ensemble methods. Overall, the study highlights the potential of deep learning to optimize resource allocation, improve web service performance, and enable data-driven decision-making in diverse domains.