Deep Learning-based Hybrid Technique for Forecasting Web Traffic

Abstract

Web traffic is a kind of time-series motion, having its highs and lows. The analysis of predicting web traffic has a greater significance for website owners, to make reliable decisions for website users. But the major gripe often faced while exploring concealed and significant details are regarding web users' different usage patterns. In this paper, we apply hybrid-based deep learning algorithms which combine two different architectures of Recurrent Neural Network (RNN), Long Short Term Memory (LSTM), and Gated Recurrent Unit (GRU). The outcome of our hybrid model is acquired by using the ensemble method of stacking. The Web Traffic Time Series Forecasting(WTTSF) dataset by Kaggle is being used to predict future traffic of Wikipedia articles. We use mean squared error, mean absolute error, and $R^{2}$ as major conventional evaluation metrics and it offers less error even though it has data randomness over a large scale.