A hybrid Bayesian network and tensor factorization approach for missing value imputation to improve breast cancer recurrence prediction

Abstract

Data mining and machine learning approaches can be used to predict breast cancer recurrence. However, real datasets often include missing values for various reasons. In this paper, a hybrid imputation method is proposed with respect to the dependency between the attributes and the type of incomplete attributes in order to especially improve the prediction of breast cancer recurrence. After splitting the dataset into two discrete and numerical subsets, first missing values of the discrete fields are imputed using Bayesian network. Then, using Tensor factorization, the integrated dataset, which comprises of the filled-subset of the previous stage and numerical missing values subset, is constructed so that both continuous missing values are imputed and the accuracy of imputation is enhanced. We evaluated the proposed method versus six imputation methods i.e. mean, Hot-deck, K-NN, Weighted K-NN, Tensor factorization and Bayesian network on three datasets and used three classifiers, namely decision tree, K-Nearest Neighbor and Support Vector Machine for recurrence prediction. Experimental results show that the proposed method has as average 0.26 prediction improvement. Also, the prediction performance of the proposed approach outperforms all other imputation-classifier pairs in terms of specificity, sensitivity and accuracy.