Aprendizagem de Classificadores para Identificação de Bactérias: relação entre as medidas de complexidade de dados e o desempenho dos classificadores

Abstract

In the agricultural environment, some bacteria have been used as active in biocontrol and plant growth. This has motivated the development of software tools to automatically detect their presence in soil samples. One way to proceed with this identification is the development of classifiers that use MALDI / TOF mass spectra patterns to check the frequency of certain ribosomal proteins in the sample. The selection of a classification function that fits the target problem has a great influence on the classifier’s performance, this has encouraged the use of scores, called data complexity measures. Such scores describe certain characteristics of the database and may provide support for choosing the classification function. During the process of generating data from mass spectrometry, it is common for data to be unbalanced, which adversely affects the data complexity measures. Considering the above, this work applies an experimental protocol to verify the influence of unbalanced data on the performance of classifiers and on complexity measures. The classifying models used in the experiments were logistic regression and QDA, which were trained to identify bacteria of the genera Bacillus and Rhizobium. The performance of the classifiers showed a strong to moderate relationship with the unbalanced data problem. Two data complexity indexes, L2B and N3B, have been proposed and submitted to tests along with the indexes found in the literature. The results show that the measures F3, Density, N3B and L2B are related to the performance of the classifiers trained with unbalanced data. Such measures were evaluated for their ability to predict the balanced accuracy of the models. When identifying bacteria of the genera Bacillus, the measure of best relation to the performance of the models was the N3B measure. In the case of the identification of the genera Rhizobium, the measure of best association with the logistic model was L2B and N3B for the quadratic model.