Competitive Coevolution for Training Neural Network Classifiers on Multi-Class Imbalanced Data - Experimental Results

The research hypothesis is that competitive coevolutionary learning is an effective method to deal with the problem of classifier learning from imbalanced data distributions. Two candidate algorithms were compared with 6 others on 12 learning problems, and the results show that competitive coevolution is an effective and computationally efficient solution. The data include all experimental results in Excel files: 10 independent runs per algorithm, on 4 different classification problems with 3 levels of class imbalance per problem. Data also include: Excel files summarising the results for each data set; the data sets (training set followed by test set in the same file), one file for the patterns (_In.txt) and one for the corresponding one-hot encoded desired classifier output (_Out.txt); and a plot of the 2D and 3D classification problems.