Selection of Niche-based Models With Minimum Description Length
... ... · Journal of Computational Interdisciplinary Sciences · 2011
Environmental issues are calling the attention of people all over the world, mainly in Brazil, which has one of the richest fauna and flora on Earth.Modeling of species geographical distribution is a technique that has been applied in many tasks related to biodiversity conservation.One of the problems of modeling species geographical distribution is to select an adequate set of environmental layers.A frequency distribution of each environmental layer can be represented by a histogram and the cut points of the histograms can be viewed as models.One of the classical problems in selecting a model is overfitting, that is, the super adjustment of the model to the observed data.The Minimum Description Length (MDL) principle has the property of avoiding overfitting when learning the parameters of the model.Thus, this is a promising strategy to be applied in the selection of any kind of model.The MDL principle searches for a model with the shortest description based on the observed data.This is done by finding regularities in data that are used to compress them.This principle was already successfully applied to probability density estimation by regular histograms.Nevertheless, there is a waste in the model representation when the data is non-uniformly distributed because of the high bin count needed to represent the details of high density data.Thus, the aim is to present how the MDL principle with irregular histograms can be used to select a good set of environmental layers.This strategy prevents the waste when representing parts of the data with low density.