Bin-BB: Binning with Branch Bound feature selection for improved diabetes classification

摘要

Neural Networks (NNs) are attractive for any classification problem. They are using extensively in the medical domain. As the medical data by nature is noisy and high-dimensional these NNs require more training time and provides poor generalization. These problems are avoided if pre-processing of data is done prior to training. Branch & Bound (B& B) based feature selection is one of the pre-processing technique extensively used to remove noisy and redundant features in the dataset. In this paper, we proposed a new criterion function based on likelihood ratio of class distributions to select more relevant features along with equal-depth binning of features. Pima Indians Diabetes dataset is chosen for experiments. Experimental results proved that proposed B&B equal-depth Binning (Bin-BB) helped in reducing the search space by reducing the height of B&B feature tree. The new criterion function helped for effective and efficient feature selection. As a result, performance of majority NN-based classifiers is improved on PID dataset.