Area-to-point kernel regression on streaming data

Abstract

Spatial data streams are often referenced to an areal spatial unit such as a polygon rather than to a precise point location. This is the case when geo-referencing is done by user IP addresses or from a mobile phone cell ID in various location-based service applications. One problem of interest in this case is spatial modelling of various spatially continuous quantities, such as an intensity of the usage of particular service in the area. This paper investigates a machine learning framework that account for area-to-point data processing. The approach is based on so-called vicinal risk minimization principle. It is elaborated in detail for a class of kernel recursive algorithms developed for distributed processing of streaming data. Concrete examples of kernel computations are provided and the method performance is investigated experimentally.