Star Ensemble:A Novel Algorithm for Spatio-Temporal Data Decomposition and Interpolation

摘要

Spatio-temporal data is becoming increasingly prevalent in our society. This has largely beenspurred on from the capability of building arrays and sensors into everyday items, along withhighly specialised measuring equipment becoming cheaper. The result of this prevalence canbe seen in the wealth of data of this kind that is now available for analysis. This spatiotemporaldata is particularly useful for contextualising events in other data sets by providingbackground information for a point in space and time. Problems arise however, when thecontextualising data and the data set of interest do not align in space and time in the exact wayneeded. This problem is becoming more common due to the precise data recorded from GPSsystems not overlapping with points of interest and not being easily generalised to a region.This is Interpolating data for the points of interest in space and time is important and a numberof methods have been proposed with varying levels of success. These methods are all lacking inusability and the models are limited by strict assumptions and constraints. This paper proposesa new method for the interpolation of points in the patio-temporal scope, based on a set ofknown points. It utilises an ensemble of models to take into account the nuanced directionaleffects in both space and time. This ensemble of models allows it to be more robust to missingvalues in the data which are common in spatio-temporal data sets due to variation inconditions across space and time. The method is inherently flexible, as it can be implementedwithout any further customisation whilst allowing for the user to input and customise their ownunderlying model based on domain knowledge. It addresses the usability issues of othermethods, accounts for directional effects and allows for full control over the interpolationprocess.