Métodos de seleção de atributos e análise de componentes principais: um estudo comparativo

Abstract

Neoplasm is a major challenge for researchers because of its high complexity. Despite advances in diagnosis, studies point out that in addition to data analysis, methods to optimize and aid the decision-making process are necessary. In this sense, the dimensionality reduction of data has contributed significantly, helping in this process, due to the large number of genes (attributes) compared to the number of samples (classes). This work, therefore, aims to provide a comparative study between two methods of dimensionality reduction, applied to three databases in the field of gene expression: LungCancer-Michigan, LungCancer-Ontario and LungCancer-Harvard, all related to lung cancer. The methods applied were: Attribute Selection and Principal Component Analysis (PCA), both used as a pre-processing step in Data Mining. The classification algorithms chosen were Naive Bayes, SVM, J48, 1-NN, 3-NN, 5-NN and 7-NN. Weka was used as a software for analyses procedures. A series of experiments was performed to evaluate the accuracy and applicability of the algorithms for both methods. As a result, significant advances in the hit rate (accuracy) of the classifiers involving the methods were evidenced, using Cross-Validation as the assessment criterion. The Wrapper approach, from the Attribute Selection method, obtained the best results for the three analyzed databases. The Principal Component Analysis method, even presenting lower hit rate, could not be ruled out. The Naive Bayes, SVM and 1-NN algorithms presented the best performance within the databases. The attributes (genes) which presented the highest frequency in the databases were denoted. Therefore, from the chosen subsets, these can be submitted to specific analyzes in order to direct more precise diagnoses.