LSTM For Web Visit Forecasting with Genetic Algorithm and Predictive Bandwidth Allocation
Aji Gautama Putrada, Nur Alamsyah, Ikke Dian Oktaviani, Mohamad Nurkamal Fauzan · 2024
Bandwidth allocation in computer networks has brought many advances, including dynamic leveraging and intelligent techniques. However, data from high-layer computer networks, such as web visit datasets, has great potential to enhance state-of-the-art bandwidth allocation. This research aims to use long short-term memory (LSTM) for web visit forecasting, which is urgent because it can significantly enhance bandwidth allocation in computer networks, leading to more efficient and intelligent network management. We utilize the “Page Loads” variable in the web visit forecasting dataset, seeing the close relationship between page loads and bandwidth allocation optimization. We performed pre-processing and analysis in the data preparation stage, including applying normalization and partial auto-correlation function (PACF). We carry out hyperparameter tuning in the LSTM model, where this step utilizes the optimization power of genetic algorithms (GA). We use metrics such as accuracy loss, LSTM model size, mean absolute error (MAE), and $r 2$ throughout the process. The test results show that our PACF analysis determines that the optimum neuron value is 4. With this number of neurons, the model size is 31.1 kB, whereas 40 neurons will result in a model size of 127.9 kB. This reduction in model size brings an accuracy loss of 2.2%. Then, the optimization results by GA provide an epoch, batch size, and learning rate of 30,20, and 0.01, respectively. Then, the GA results suggest using the full sequence output, where the best fitness function is $\mathbf{0. 0 5 9}$. Finally, the results of LSTM training with hyperparameter tuning from GA give a loss value of 0.042 and $r 20.913$. Then, the model’s loss and $r 2$ values in testing are 0.043 and 0.914, respectively. The novelty of this research is the LSTM model for web visit forecasting with the “Load Page” variable.