Improving neural-based classification of databases with overlapped classes: The case of star/galaxy segregation

There are many real-life classification problems where class overlapping severely limits the classification accuracy. In these situations is difficult to build automatic classifiers that obtain good generalization performance. An interesting case is found in the separation of stars and galaxies, which arises in galactic or extragalactic studies. There are many astronomical analysis packages which deal with this problem; for example, the very popular package SExtractor (Source Extractor) has incorporated a multi-layer perceptron (MLP) neural network classifier. We believe that SExtractor performance is suitable for improvement. In our way for building a better classifier, we analyzed the behavior of MLP-based classifiers for this kind of data. In this paper we present an experiment where, using WEKA, we have automatically selected the best characteristics to discriminate galaxies from stars and automatically selected the topology of a MLP that best defined the decision region. Our classifier obtained slightly better results than SExtractor when compared to classifications obtained by a human expert, using less computational resources that SExtractor. However, we conclude that more specific information about the problem needs to be used to build a better separator of star/galaxies.