Orlando Ohashi is a Professor at UFRA, Brazil. He completed is MAPi PhD program at University of Porto in 2012 under the supervision of Prof. Luís Torgo

PhD Thesis

Title: Spatio-Temporal Prediction Methods

Supervisor: Luis Torgo;

MAP-i Doctoral Programme in Computer Science

Finished: Dec/2012

Abstract

The volume of data that is currently collected and stored was unthinkable a few years ago. This amount of information makes data and all phases of the process of collecting, storing and making sense of it, extremely important. Both academia and industry are working in this process. Data mining is a key component to help users to make sense of this huge amount of data. This research field includes a large set of tasks. This thesis addresses the problem of prediction using spatio-temporal data, i.e. data that are indexed both in time and in space.

The work presented in this thesis is driven by several real world applications: (i) monitoring and controlling water quality parameters within the water distribution network at Porto, Portugal; (ii) forecasting water consumption for a water distribution company in Spain; (iii) forecasting wind speed in some wind farm in the US; and (iv) filling in missing pixels of images.

Our work is organized in an incremental fashion by addressing different particularities of our applications. Concretely, we first address temporal prediction problems, then spatial prediction tasks and finally we focus on spatio-temporal data sets.

For temporal data we propose a new class of forecasting tasks that we name 2D-interval predictions, which consists on trying to obtain a forecast of the expected range of values for a future time interval. We formalize this task, propose a solution to it and establish the correct way of evaluating models for these tasks. Our extensive experimental tests show the advantage of our proposal for these tasks.

Regards, spatial data we address the problem of spatial interpolation by proposing a new methodology based on two key ideas: (i) transforming the problem into a regression task and (ii) describing the spatial dynamics by spatial indicators. This methodology differentiates itself from the state of the art in that it allows the use of data from nonnearby regions to forecast the value for a certain location, thus somehow contradicting the first law of geography. We have extensively evaluated this methodology in problems of filling in missing pixels in photos. Our results show a clear advantage of our proposal when compared to the state of the art in spatial interpolation.

Finally, we proposed a new technique to improve the prediction accuracy in spatio-temporal data. Our technique differs from the most common approaches, in that is uses spatiotemporal properties of the data to improve the predictive accuracy. Namely, we propose a series of spatio-temporal indicators whose goal is to describe the spatio-temporal dynamics of the data for each location. We extensively test our technique using real world wind speed data, and we observed a clear advantage of our proposal when compared to several alternative methods that can be applied to these problems.