Solving multiclass classification problems using combining complementary neural networks and error-correcting output codes

Abstract

This paper presented an innovative method, combining Complementary Neural Networks (CMTNN) and Error-Correcting Output Codes (ECOC), to solve multiclass classification problem. CMTNN consist of truth neural network and falsity neural network created based on truth and falsity information, respectively. Two forms of ECOC, exhaustive code and random ECOC, are considered to deal with k-class classification problem. Exhaustive code is applied to the problem with 3 ≤ k ≤ 7 whereas random ECOC is used for k > 7. In the experiment, we deal with feed-forward backpropagation neural networks, trained using 10 fold cross-validation method and classified based on two decoding techniques: minimum distance and T > F. The proposed approach has been tested with six benchmark problems: balance, vehicle, nursery, Ecoli, yeast and vowel from the UCI machine learning repository. Three data sets: balance, vehicle and nursery are dealt with exhaustive code while random ECOC is applied for Ecoli, yeast and vowel. It was found that our approach provides better performance compared to the existing techniques considering on either CMTNN or ECOC.