Hybrid associative memories for imbalanced data classification: An experimental study

Abstract

Hybrid associative memories are based on the combination of two well-known associative networks, the lernmatrix and the linear associator, with the aim of taking advantage of their merits and overcoming their limitations. While these models have extensively been applied to information retrieval problems, they have not been properly studied in the framework of classification and even less with imbalanced data. Accordingly, this work intends to give a comprehensive response to some issues regarding imbalanced data classification: (i) Are the hybrid associative models suitable for dealing with this sort of data and, (ii) Does the degree of imbalance affect the performance of these neural classifiers Experiments on real-world data sets demonstrate that independently of the imbalance ratio, the hybrid associative memories perform poorly in terms of area under the ROC curve, but the hybrid associative classifier with translation appears to be the best solution when assessing the true positive rate. © 2013 Springer-Verlag Berlin Heidelberg.