Clustering schemes

• Distance-based: assumes metric space on attributes
• Easier on numerical data since basic notions are natural
• Distance metric easier to formulate (e.g. weighted Euclidean)
• “Distance” from a reference point
• Quantities such as averages, correlations, centroids, and so forth are well-defined.
• Clustering on categorical data is generally harder since metric notions have to be defined.
• Model-based:
• estimate a density (e.g. KDE, mixture of gaussians, …)
• go bump-hunting
• compute P(X_i|Cluster j)
• Partition-based:
• enumerate partitions and score each: usually using some adhoc scoring scheme (e.g. conceptual clustering in AI)

---