Novel Framework for Improving the Desired Structure Prediction on Imbalance Data Set

Abstract

Desired target prediction has important significance in rational synthesis of materials. In this paper, taken (6,12)-ring-containing structure as the desired target, a novel framework is proposed for improving the prediction performance on the microporous aluminophosphates (AlPOs) database. In contrast to existing data processing techniques for class imbalance problem, the proposed framework first explores the intra-class distribution of majority class by clustering such that multiple specific models can be established according to the distribution. The main advantage is that one model can characterize the clustered data well. Then, Q times n-fold cross-validation procedure is applied to assess the prediction performance. Finally, we compare the proposed framework with two existing data processing procedures. The comparison results demonstrate that the desired target predictions can achieve improved performance remarkably.