Integrative Unsupervised and Supervised Learning Approaches for Breast Cancer Subtype Classification Using Gene Expression Data

Breast cancer is a heterogeneous disease with distinct molecular subtypes that require precise classification for personalized treatment strategies. This study proposes an integrative methodology combining unsupervised and supervised learning techniques (hybrid learning) to classify breast cancer subtypes using gene expression data from the Gene Expression Omnibus (GEO) repository. Hierarchical clustering is employed as an exploratory unsupervised approach, using both Euclidean distance and Pearson correlation to reveal intrinsic data structures. For supervised classification, four machine learning models—Logistic Regression, Support Vector Machine (SVM), Random Forest, and Multilayer Perceptron (MLP)—are applied. These models are further optimized via the Optuna framework to enhance performance through hyperparameter tuning. SHAP values are used to assess the importance of features, contributing to model interpretability. The results show that supervised and unsupervised approaches are complementary, offering both accuracy and insight into subtype differentiation. Notably, models optimized by Optuna significantly outperformed non-optimized counterparts. The findings emphasize the potential of combined methodologies in supporting early and accurate diagnosis of breast cancer subtypes.