We developed a method that properly deals with autocorrelation in data that are not independently and identically distributed (i.i.d.) and provides a multi-level insight into the autocorrelation phenomenon. The method is based on the concept of predictive clustering trees (PCTs) and works for different predictive modeling tasks, including classification and regression, as well as some clustering tasks. We applied this method to several real world problems of spatial regression and classification, as well as problems of network regression coming from the areas of social and spatial networks. STOJANOVA, Daniela, CECI, Michelangelo, APPICE, Annalisa, MALERBA, Donato, DŽEROSKI, Sašo. Global and local spatial autocorrelation in predictive clustering trees. Lect. notes comput. sci., 2011, vol. 6926, str. 307-322. [COBISS.SI-ID 25200423]
COBISS.SI-ID: 25204775
We propose the use of random forests and bagging of predictive clustering trees in the domain of medical image annotation with labels organized into a hierarchy. The experiments show that ensembles of predictive clustering trees perform consistently better than SVMs. Second, SIFT descriptors are the most discriminative. Next, combinations of several descriptors improve the predictive performance of the classifiers. Finally, the results of the annotation of this image database are the best results reported so far both in the literature and at image annotation competitions.
COBISS.SI-ID: 24848423