Yapay öğrenme ile uzam-zamansal modelleme

An ideal machine learning algorithm for spatiotemporal modeling should be able (i) to integrate both temporal and spatial data from different sources, (ii) to discover patterns and, (iii) to make inference, without human intervention. Gaussian processes provide a Bayesian framework for analyzing spatiotemporal data, which were used widely to estimate values across space and time, yet their computational and storage complexity have been a limiting factor when it comes to application. In this thesis, we proposed computational frameworks that integrate Gaussian processes into spatiotemporal modeling scenarios with a particular focus on scalable inference by exploiting the structure of the covariance matrix generated by matrix multiplication of spatial and temporal covariance matrices. We also aimed to increase the interpretability using kernel methods, which have deep connections with spatial statistics. With the combination of multiple kernel learning and structured Gaussian processes, we increased both accuracy and expressiveness of the inference. We showed the power of these methods on real-world regression data sets. Our proposed methods were applied to a spatiotemporal data set of a vector-borne disease using official patient records in Turkey. We showed our proposed machine learning algorithms were better than their counterparts in terms of accuracy. In addition, our developed methods are also more interpretable, which means that they are able to answer questions drawn from the domain of public health and give insight to policy makers for quick response planning and resource allocation.