Mariana Oliveira is a PhD student at the MAPi Doctoral Programme in Computer Science, hosted at Faculty of Sciences, University of Porto (FCUP) and the Laboratory of Artificial Intelligence and Decision Support (LIAAD – INESC TEC). Her research currently focuses on predictive analytics for dependent data, supported by an FCT-MAPi grant. Her advisors are Prof. Luís Torgo and Prof. Vítor Santos Costa.

MSc Thesis

Title: Propositional and Relational Approaches to Spatio-Temporal Data Analysis

Supervisor: Luis Torgo; Co-supervisor: Vitor S. Costa, University of Porto, Portugal

MSc in Computer Science, FCUP-U.Porto

Finished: Oct/2015


Understanding spatio-temporal phenomena is a fundamental challenge in the field of Data Mining with applications ranging over a myriad of domains. One such challenge arises from spatio-temporal databases of time-varying data that can be represented by an evolving thematic map. Several approaches aiming at describing spatio-temporal data and predicting future values have already been proposed. Among them, we find propositional approaches that work on a single table, and relational approaches with the ability to work on multiple related tables. We review both types of approaches to association rule learning and regression problems, and enumerate the challenges faced.

Our motivating application concerns wildfires in Portugal, which every year have a strong socio-economical and environmental impact in the country. We adapt a notion of spatio-temporal neighbourhood to include spatial direction, propose a concept of simplified border for heterogeneous spatial objects, build spatio-temporal indicators based on these notions, design relational predicates that deal with numerical attributes and include the temporal and spatial dimensions, and deploy a re-sampling tecnhique to improve regression under an imbalanced domain. We apply a relational and a propositional approach to the problems of understanding and predicting wildfires in mainland Portugal, and draw comparisons between the two. We are able to find strong association rules and accurately predict the yearly percentage of burnt area in each Portuguese civil parish in spite of the several challenges posed by this problem.


  • Forescasting with Dependent Data
  • Spatiotemporal data analysis
  • Forecasting