ICT: Iterative Clustering with Training: Preliminary Results

We introduce Iterative Clustering with Training (ICT), a hybrid approach that combines clustering with iterative classifier-based refinement. Unlike traditional clustering methods, ICT leverages supervised learning to reassign ambiguous boundary points, improving cluster quality. The proposed approach starts with an initial clustering phase, followed by the computation of silhouette scores to separate high-confidence core points from uncertain boundary points. A classifier (e.g., Random Forest) is then trained on core points to predict labels for boundary points, refining clusters in an iterative process until convergence is achieved. To avoid trivial clusters, minimum cluster ratio constraints are implemented. Experimental results on multiple benchmark datasets show that ICT outperforms traditional clustering methods such as k-means, HDBSCAN, and spectral clustering across several evaluation metrics, such as Adjusted Rand Index, Normalized Mutual Information, and Accuracy. ICT particularly excels at handling clusters that overlap, have noise, and complex geometries, where standard methods often fail. Key advantages include adaptive core selection, iterative refinement, and classifier flexibility, making ICT suitable for applications in customer segmentation, anomaly detection, and biological data analysis.