Duplicate-sampling of difficult-to-classify sheme

Abstract

The multilayer perceptron (MLP) network is applied successfully in solving many pattern classification problems. The famous learning algorithms used to train the MLP network is the backpropagation (BP) algorithm. One disadvantage regarding to the BP algorithm is its lengthy computational time. Therefore, a number of techniques are developed to handle this situation. Most of those techniques are still based on a concept of improving the convergence of the network’s error. In this thesis, we present another view, based on distribution of data and information gain, to solve the time-consuming problem called duplicate-sampling of difficult-to-classify scheme. The proposed technique is designed as an enhancement tool to apply with the standard BP algorithm. The technique utilizes the information gain to split data space into subspaces based on binary splitting of numeric attribute. Data located in the impure subspaces are identified as difficult-to-classify data which are duplicated before starting the BP algorithm. During the learning process, the difficult-to-classify data will be emphasized by the BP learning. The BP learning with this technique requires less computational time to achieve the same or higher accuracy. The experiments performed on eight data sets taken from the UCI repository indicate a computational time improvement. Seven out of eight data sets showed a better result