Classification of Incomplete Data using Augmented MLP

Abstract

We introduce a new way to train a Multi-Layer Perceptron (MLP) to classify incomplete data. To achieve this, we train an MLP using a two-phased approach. In the first phase, we train an MLP using complete data. We create an augmented dataset before the second phase of training. For this, we use non-missing data, delete each feature once, and then fill it using some predefined points. After that, in the second phase, we retrain the network using the augmented dataset. The aim of this type of training is to predict the class label of an incomplete dataset. At the time of testing, when a feature vector with a missing value appears, we initially impute it using the predefined points and find the class label of the feature vector using the trained network. We compare the proposed method with an original MLP on twelve datasets using four imputation strategies. The proposed method’s performance is better compared to the originally trained MLP.