An empirical comparison of predictive models for web page performance

Context: The quality of user experience is the cornerstone of any organization's successful digital transformation journey. Web pages are the main touchpoint for users to access services in a digital mode. Web page performance is a key determinant of the quality of user experience. The negative impact of poor web page performance on the productivity, profits, and brand value of an organization is well-recognized. The use of realistic prediction models for predicting page load time at the early stages of development can help minimize the effort and cost arising out of fixing performance defects late in the lifecycle. Objective: We present a comprehensive evaluation of models based on 18 widely used machine learning techniques on their capability to predict page load times. The models use only those metrics which relate to the form and structure of a page because such metrics are easy to ascertain during the early stages with minimal effort. Method: The machine learning techniques are trained on more than 8,700 pages from HTTP Archive data, a database of web performance information widely used to conduct web performance research. The trained models are then validated using the 10-fold cross-validation method and accuracy measures like the Pearson correlation coefficient (r), Root Mean Square Error (RMSE), and Normalized Root Mean Square Error (NRMSE) are reported. Results: Radial Basis Function regression and Random Forest outperform all other techniques. The value of r ranges from 0.69-0.92, indicating a high correlation between the observed and predicted values. The NRMSE varies between 0.11-0.16, implying that RMSE is less than 16% of the range of actual value. The RMSE improves by 41%-54% compared to the best baseline prediction model. Conclusion: It is possible to build realistic prediction models using machine learning techniques that can be used by practitioners during the early stages of development with minimal effort.