An imbalanced classification problem arises when the distribution of samples across classes is uneven or skewed. Managing small and imbalanced training datasets remains a significant obstacle in machine learning, particularly in fields like bioinformatics and medical research. Such imbalances can produce biased models that perform poorly on minority classes, over-rely on certain features, and fail to identify the true underlying patterns in the data. This study introduces a feature selection method grounded in gene connectivity, combined with a class-balancing strategy, to construct a machine learning model from imbalanced gene expression data. The approach was evaluated on a rheumatoid arthritis dataset containing 28 normal samples and 152 rheumatic samples. Using weighted gene co-expression network analysis (WGCNA), the original 27,991 features were reduced to 601. These selected features served as input for machine learning classification models trained first on the imbalanced data and then on balanced data generated via the Spread Sub-Sample technique. Results showed that two classifiers achieved higher accuracy on the imbalanced dataset compared to the balanced version, highlighting the tendency of many classifiers to become biased when trained on imbalanced data. Logistic regression attained an improved accuracy of 95%, while the decision tree and IBK algorithms recorded reduced accuracies of 81% and 91%, respectively. In conclusion, combining feature selection with class-balancing techniques is essential to reduce model execution time and improve accuracy, especially when working with RNA-Seq gene expression data.
