Performance Evaluation of Classification Pipelines Builded with Restricted Boltzmann Machine and Several Classifiers

Many classification methods in the literature are used to achieve the highest classification accuracy. The restricted Boltzmann machine is an artificial neural network with unsupervised learning, however it is gained importance as a learning component to extract attributes. In our study, we compared the performances of five different pipelines builded with the Bernoulli-type restricted Boltzmann machine and different classifiers. Experiments have been carried out on these pipelines by using logistic regression, decision tree, Gaussian naive Bayes, Ada Boost, and random forest classifier, respectively. The classification resultant changes were observed through the use of pipelines compared to the stand alone classification results obtained by these classifiers. Experimental results were obtained with the use of MNIST handwriting digit recognition data set. In these experiments, two different orders of hyper parameters of the restricted Boltzmann machine were used. According to these results, it was seen that the classification accuracy of the stand alone classifier became better through the use of pipelines in the experiments. The highest performance was achieved with a classification success rate of 97.19%. Models using the proposed pipeline design have improved the average performance of related invidividual classifiers from at least 1% to at most 33%. Discussions are also included