Confusion matrices for improving performance of feature pattern classifier systems
Ignas Kukenys, Will Neil Browne, Mengjie Zhang · 2011
Learning Classifier Systems (LCS) have not been widely applied to image recognition tasks due to the very large search space of pixel data. Assimilating the image domain's Haar-like features into the XCS framework, the feature pattern classifier system (FPCS) has produced promising results in the numeral recognition task. However for large multi-class image classification problems the training rates can be unacceptably slow, whilst performance does not match supervised learning approaches. This is partially due to the fact that traditional LCS only retain limited information about the problem examples. Confusion Matrices show the classes that a learning technique has difficulty separating, but require supervised knowledge. This paper shows that the knowledge in a confusion matrix is beneficial in directing learning. Most importantly the work shows that confusion matrices can be beneficially adapted to non-supervisory learning.