Analyzing and Improving Prediction of Spatiotemporal Signal Data Using Grid Search on Graph Convolutional Networks
Seyyed Ali Mohammadiyeh, Behzad Soleimani Neysiani · 2023
Spatiotemporal signal processing is one of the complex and hot topics, especially in web mining like web traffic analysis. The web pages and their links are a graph, and their content (e.g., visits) can be a signal. The PyTorch Geometric Temporal is introduced for spatiotemporal signal mining. This study analyzes spatiotemporal signals like visits of Wikipedia mathematics pages using the PyTorch Geometric Temporal library to improve their visit prediction during the time using a grid search for hyper-parameter adjustment and analyzing the effect of each parameter. The results for five datasets including Pedal Me, Chickenpox, Wikipedia Maths, England Covid and Monte Video Bus show more than 8.03% relative improvement for the GConvGRU algorithm versus basic related work in state-of- the-art based on about 129,000 experiments. Besides, it should be considered that lags and node feature parameters must be the same, and lower learning rate and epochs, and higher training ratio and filter size are the best possible values.