Advances in Machine Learning (ML) present new opportunities to systematically analyze the spatial complexity of urban form. This study presents a proof-of-concept for an interpretable methodological framework for clustering urban typologies. The methodology employs the Getis-Ord Gi* spatial autocorrelation metric as positional information to encourage the creation of spatially homogenous clusters. Clustering is performed using UMAP, a non-linear dimensionality reduction algorithm along with BIRCH, a scalable unsupervised clustering algorithm. The method utilizes 17 morphological indicators that capture urban form attributes at the block, plot and building scale. The proposed framework is pilot tested on the metropolitan area of Thessaloniki, Greece, revealing 14 distinct urban typologies that are organized into 5 families with similar characteristics. The typologies reveal, in an almost Conzenian fashion, patterns of urban development that are rooted in the city’s modern history. Results are validated both quantitatively with performance indicators and qualitatively using aerial imagery and established knowledge on Thessaloniki’s urban planning and evolution.