Date of Award


Document Type


Degree Name

Master of Science in Computer Science


School of Computer Science and Engineering

First Reader/Committee Chair

Zhang, Yan


Data collected from the real world is often imbalanced, meaning that the distribution of data across known classes is biased or skewed. When using machine learning classification models on such imbalanced data, predictive performance tends to be lower because these models are designed with the assumption of balanced classes or a relatively equal number of instances for each class. To address this issue, we employ data preprocessing techniques such as SMOTE (Synthetic Minority Oversampling Technique) for oversampling data and random undersampling for undersampling data on unbalanced datasets. Once the dataset is balanced, genetic programming is utilized for feature selection to enhance performance and efficiency. For this experiment, we consider an imbalanced bank marketing dataset from the UCI Machine Learning Repository. To assess the effectiveness of the technique, it is implemented on four different classification algorithms: Decision Tree, Logistic Regression, KNN (K-Nearest Neighbors), and SVM (Support Vector Machines). Various metrics including accuracy, balanced accuracy, recall, F-score, ROC (Receiver Operating Characteristics) curve, and PR (Precision-Recall) curve are compared for unbalanced data, oversampled data, undersampled data, and cleaned data with Tomek-Links for each algorithm. The results indicate that all four algorithms perform better when oversampling the minority class to half of the majority class and undersampling the majority class examples to match the minority class, followed by performing Tomek-Links on the balanced dataset.

Included in

Data Science Commons