Assessing Wireless Data Services with Machine Learning and Geostatistics

The California Public Utilities Commission collects and publishes data on mobile broadband service from major wireless service providers. The data is collected twice yearly at over 1900 measurement locations across California. The data can be used to assess the "digital divide", to track improvements in wireless service over time, and to help consumers select a service provider. For example, a consumer might want to know which provider offers the best service quality at important locations, like home and work. Here we apply machine learning and geospatial estimation methods to estimate each provider's wireless service quality at arbitrary locations across the state. We focus on TCP download performance. We find that Kriging, which takes into account the spatial correlation of service quality across measurement locations, yields better predictions than k-nearest neighbor and inverse distance weighting. We observe similar results when we rank providers rather than estimate the quality of service they provide. Finally, because average service performance across the state can be used to rank service providers, we estimate the minimum number of measurement locations that could be used to reliable rank service provider performance.