Unsupervised Machine Learning: Datasets Without Outcomes

Abstract

Often either a dataset doesn’t contain an obvious “outcome” or we wish to explore the entire data set to see if there is some natural “order” to the data. In such cases, unsupervised machine learning is appropriate. Once again, unsupervised means that there isn’t an outcome to compare the results of a model to. It is tempting to try to “force” a regression or classification model, but often it is quite enlightening to use unsupervised methods to better understand the dataset. If the data are nominal “marketbasket” lists of “transactions” (for example the set of laboratory tests ordered at one time for a particular patient on a particular day; or items purchased at the supermarket), the technique of association analysis is most appropriate. If the data are quantitative, with a metric of proximity available, a clustering technique can be used.