Hierarchical Clustering-Based Synthetic Minority Data Generation for Handling Imbalanced Dataset

Predictive modeling is a new area of data science and machine learning that is gaining popularity. It provides sustained business growth, accurate future predictions, and trend estimations. Predictive modeling is the process of creating, processing, and validating a model that may be used to make future predictions using known results. Predictive modeling depends on the complete and precise datasets, however some of the datasets are imbalanced in nature that leads to data misclassification. Models trained on an imbalanced dataset with a small number of minority class instances, despite their high accuracy, would perform poorly during training. In this paper, an approach for synthetic minority class data generation using agglomerative hierarchical clustering and ward’s linkage criteria is proposed. Experimentation is carried out using five real-world datasets, namely Abalone, Page Blocks, Pima, Vehicle, and Yeast available at KEEL Data Repository. Testing of the experimentation is done using the SVM classifier with radial-bias kernel function. Data visualization is performed for understanding statistical properties, correlations, and distribution of class instances in the feature space. The classifier model is evaluated for before and after synthetic data generation using f-measure, recall, precision, and accuracy.