Unsupervised Learning for customer segmentation: comparing advanced clustering methods in practice What changes when K-Means, K-Medoids, Ward hierarchical clustering, DBSCAN and Gaussian Mixtures see the same 8,950 credit-card customers? Dataset source:...
More
Unsupervised Learning for customer segmentation: comparing advanced clustering methods in practice What changes when K-Means, K-Medoids, Ward hierarchical clustering, DBSCAN and Gaussian Mixtures see the same 8,950 credit-card customers? Dataset source: Arjun Bhasin — Credit Card Dataset for Clustering, Kaggle (CC0: Public Domain). The Kaggle data card describes roughly 9,000 active credit-card holders observed over the previous six months, with 18 customer-level behavioral variables. The local file used in this exercise contains 8,950 rows and the same 18 columns. The question behind the exercise If I keep the data, preprocessing and dimensionality-reduction space fixed, how much of the final segmentation is determined by the clustering algorithm itself? Most introductory clustering examples end with one model and one colorful scatter plot. I find this useful for learning the API, but it tends to hide the most difficult part of unsupervised learning: there is no target label telling
Less
