A hybrid meta-cognitive restricted boltzmann machine classifier for credit scoring

Meta-cognition with self-regulated learning equips a machine learning algorithm to make judicious decisions about every sample in the training data set. Due to this capability, meta-cognitive machine learning algorithms exhibit better generalization behavior. In the past, numerous works have focused on studying the effect on meta-cognition for learning patterns from data in a supervised fashion. In this paper, we attempt to study the effect of meta-cognition towards efficient feature representation in a Restricted Boltzmann Machine (RBM). To this end, we develop a Meta-cognitive Restricted Boltzmann Machine (McRBM) that decides what-to-represent, how-to-represent and when-to-represent for each sample in the training data set through its Sample deletion, Feature Representation and Sample Reserve strategies, respectively. The meta-cognitive component of McRBM helps to improve the generative training of RBM. The features thus generated by RBM are then used in classification through back propagation learning, which is derived based on the hinge-loss error function. The McRBM is used to solve credit scoring problem using the German credit data set, Australian credit data set and the KAGGLE credit data set. Performance of McRBM is compared against the Support Vector Machines (SVM), Extreme Learning Machines (ELM), Multi-layer Perceptron with Back Propagation (MLP-BP), and Classification Restricted Boltzmann Machine (ClassRBM). Performance results show superior classification ability of McRBM.