PATTERN TREE SPATIAL MODELS FOR ECOLOGICAL CLASSIFICATION

Abstract

This paper further extends pattern trees membership functions, by implementing a modified sigmoid distribution. In this work we use this algorithm to extract knowledge for ecological classification task from the diatoms community measured dataset, which according the biological experts are used as bio-indicators in many water ecosystem environments. The first part of the algorithm transforms the input set from crisp values into fuzzy values, and then continues the induction of the tree. The transformation is achieved by using different membership functions, which have different shape and mathematical description. This is very important because later in the induction phase this will have effect on the classification accuracy and complexity of the obtained model. The modified sigmoid function that is put on test, have several advantage over the triangular and trapezoidal functions. The experiments on diatoms classification datasets showed that sigmoid shaped function algorithm models outperform the pattern tree models build based on the trapezoidal, triangular or Gaussian MF in terms of prediction accuracy. The diatom models based on this method produced valid and useful knowledge that later in the paper is interpreted. Finally, evaluation performance analyses of the build pattern trees with classical classification algorithms is presented and discussed.