Minimum Information Trees for High Dimensional Data Visualization in Clustering

Visualizing the qualitative outcomes of clustering algorithms in high-dimensional spaces remains a persistent challenge in data analysis and machine learning. Traditional dimensionality reduction techniques often distort the underlying structure of clusters or fail to provide interpretable representations of inter-cluster relationships. In this paper, we introduce Minimum Information Trees (MINFO Trees), an information-theoretic, graph-based method for visualizing high-dimensional data with an emphasis on preserving clustering structure. By leveraging pairwise information measures and constructing information-theoretic based k-NN graphs, MINFO Trees generate data visualizations that reflect both local cohesion and global separation among clusters. Our method provides interpretable and faithful representations of clustering results, enabling qualitative evaluation of cluster quality and relationships. Experimental results on real-world datasets highlight the differences between MINFO Trees over state-of-the-art dimensionality reduction methods such as t-SNE, UMAP and PaCMAP in terms of preserving cluster topology and enhancing visual interpretability.