GWR-STARMA:A Spatiotemporal-aware Hybrid Prediction Model For Response Time Of Web Map Services By Integrating GWR And STARMA

Currently,a huge number of geospatial web services are available over the Internet for public use providing abundant geospatial resources but varying in functions and quality.In this context,discovering and selecting an appropriate geospatial web service with satisfactory quality for a particular application becomes a challenging problem;even among a group of services with same functionality.As a non-functional attribute,response time is an important measurement of web service performance.Response time impacts user experience significantly and thus plays an important role in web service selection.However,response time has significantly uncertainty and hard to predict because it is not only determined by software and hardware performance,but also affected by spatiotemporal distribution of users.So,it is difficult to predict the response time reliably.Service performance is dynamic for users from different regions at different times.OGC Web Map Service(WMS)was tooken as an example and a global distributed monitoring system was developed for collecting the moniroring data of response times for thousands of WMSs.Analysing the correlation between WMS response time and spatiotemporal factors,we found that there is positive relationship between response time and distance from user to service and response time series is stationary.What's more,neighboring users tend to similar response time series.Based on data analysis,we proposed a hybrid spatiotemporal prediction model that takes the impact of both space and time disparities of user access into account,by integrating geographical weighted regression(GWR)and spatial-temporal auto regressive and moving average(STARMA).Specifically,GWR component simulates macro-level spatiotemporal trends of response time and STARMA component captures local stochastic variations in spatiotemporal series.By comparing the predicated data with the monitoring data from different monitoring sites and different times,the feasibility was proved.The experiment result shows that the proposed model has significant predication accuracy improvement than the classical average model(AVG)and slight accuracy improvement than GWR.The GWR-STARMA model is appropriate for real-time prediction of response time with frequent fluctuations.