Classification of Incomplete Data using Augmented Dataset

Abstract We introduced a new way to train a classifier to classify different data specifically incomplete data. Here, we train a classifier using a two-phase approach. In the first phase, we train a classifier using complete data. Then, we create a new data set (augmented dataset) before the second phase of training. After that, in the second phase, we retrain the classifier using the newly created data set. At the time of testing, if a feature vector with a missing value appears, we initially impute it using various strategies and complete it. Then we try to find the class label of each complete feature vector using the trained classifier. If a clas-sifier is trained using the augmented dataset, the performance of that classifier is dependent on that augmented dataset. To overcome this issue, later, we proposed two ways to combine different augmented datasets. We compare the proposed method using five classifiers on twelve datasets. We also use four types of imputation strategies for initial imputation if a dataset has missing values. The proposed method’s performance is better compared to the originally trained classifier.