Investigating Accuracy and Diversity in Heterogeneous Ensembles for Breast Cancer Classification

Abstract

Breast Cancer (BC) is one of the most common forms of cancer among women. Detecting and accurately diagnosing breast cancer at an early phase increase the chances of women’s survival. For this purpose, various single classification techniques have been investigated to diagnosis BC. Nevertheless, none of them proved to be accurate in all circumstances. Recently, a promising approach called ensemble classifiers have been widely used to assist physicians accurately diagnose BC. Ensemble classifiers consist on combining a set of single classifiers by means of an aggregation layer. The literature in general shows that ensemble techniques outperformed single ones when ensemble members are accurate (i.e. have the lowest percentage error) and diverse (i.e. the single classifiers make uncorrelated errors on new instances). Hence, selecting ensemble members is often a crucial task since it can lead to the opposite: single techniques outperformed their ensemble. This paper evaluates and compares ensemble members’ selection based on accuracy and diversity with ensemble members’ selection based on accuracy only. A comparison with ensembles without member selection was also performed. Ensemble performance was assessed in terms of accuracy, F1-score. Q statistics diversity measure was used to calculate the classifiers diversity. The experiments were carried out on three well-known BC datasets available from online repositories. Seven single classifiers were used in our experiments. Skott Knott test and Borda Count voting system were used to assess the significance of the performance differences and rank ensembles according to theirs performances. The findings of this study suggest that: (1) Investigating both accuracy and diversity to select ensemble members often led to better performance, and (2) In general, selecting ensemble members using accuracy and/or diversity led to better ensemble performance than constructing ensembles without members’ selection.