Comparative Analysis of Feature Engineering-Based Learning and Deep Learning Approaches for Classification Problems on Imbalanced Datasets
Rajasekharan Rajasree, R. Geetha, S. Vidhya, M. A. Mukunthan, S. Hari Kumar, Jose Anand A. · 2025
This study aimed to compare the overall performance of two prominent machine learning approaches for tackling classification problems: feature engineering-based learning and deep learning. The feature engineering-based approach involved identifying and calculating relevant features, while the deep learning approach utilized raw signals without additional preprocessing. The comparative analysis was conducted on a case study dataset with imbalanced classes. For the feature engineering-based approach, the best-performing model was the XGBoost classifier trained on randomly oversampled data, achieving an average F1 score of 77%. An alternative XGBoost model, adjusted to account for the problem’s cost sensitivity, yielded a slightly lower average F1 score of 75%. In contrast, the deep learning approach achieved superior results using a 1D Convolutional Neural Network (CNN), which obtained an average F1 score of 80%. The findings suggest that the features automatically extracted by the convolutional layers in the CNN held greater relevance for classification than those manually calculated using expert knowledge.