MBCGP-FE: A modified balanced cartesian genetic programming feature extractor

Many data sets are represented by low-level or primitive features. This makes it difficult to discover relevant information via learning algorithm. Changing the way primitive data is represented can be advantageous. This can be performed using data preprocessing algorithms. A successful preprocessing algorithm should be capable of revealing the relationships among features to improve learners. These hidden relations among features can make the relevancy of the aspects of the data opaque to the learner. Automatic feature extraction is a solution to overcome this problem. This article introduces a Modified Balanced Cartesian Genetic Programming Feature Extractor (MBCGP-FE) for transforming the feature space to a smaller one composed of highly informative features through modifying the representation and operators of Balanced Cartesian Genetic Programming (BCGP). The new feature space is composed from original relevant and new constructed features which are created by discovering and compacting hidden relations among features. The size of the new feature space is determined during the optimization process. Experimental results on real data sets show that the MBCGP-FE improves the performance of learners and it is effective in reducing the dimension of data sets through the construction of new informative features. In addition, obtained results indicate the effectiveness of our proposed method in comparison with other feature extraction methods.