Clustering is a multivariate technique that groups together observations that share similar values across a number of variables. Typically, observations are not scattered evenly through p-dimensional space, where p is the number of variables. Instead, the observations form clumps, or clusters. Identifying these clusters provides you with a deeper understanding of your data.
Note: JMP also provides a platform that enables you to cluster variables. See Cluster Variables.
JMP provides four platforms that you can use to cluster observations:
• Hierarchical Cluster is useful for small and large data tables and allows character data. Hierarchical clustering combines rows in a hierarchical sequence that is portrayed as a tree. You can choose the number of clusters that is most appropriate for your data after the tree is built. See Hierarchical Cluster.
• K Means Cluster is appropriate for larger tables with up to millions of rows and allows only numerical data. You need to specify the number of clusters, k, in advance. The algorithm guesses at cluster seed points. It then conducts an iterative process of alternately assigning points to clusters and recalculating cluster centers. See K Means Cluster.
• Normal Mixtures is appropriate when your data come from a mixture of multivariate normal distributions that might overlap and allows only numerical data. For situations where you have multivariate outliers, you can use an outlier cluster with an assumed uniform distribution. See Normal Mixtures.
You need to specify the number of clusters in advance. Maximum likelihood is used to estimate the mixture proportions and the means, standard deviations, and correlations jointly. Each point is assigned a probability of being in each group. The EM algorithm is used to obtain estimates.
• Latent Class Analysis is appropriate when most of your variables are categorical. You need to specify the number of clusters in advance. The algorithm fits a model that assumes a multinomial mixture distribution. A maximum likelihood estimate of cluster membership is calculated for each observation. An observation is classified into the cluster for which its probability of membership is the largest. See Latent Class Analysis.
|
Method |
Data Type or Modeling Type |
Data Table Size |
Specify Number of Clusters |
|---|---|---|---|
|
Hierarchical Cluster |
Any |
With Hybrid Ward, up to hundreds of thousands of rows With Fast Ward, up to 200,000 rows With other methods, up to 5,000 rows |
No |
|
K Means Cluster |
Numeric |
Up to millions of rows |
Yes |
|
Normal Mixtures |
Numeric |
Any size |
Yes |
|
Latent Class Analysis |
Nominal or Ordinal |
Any size |
Yes |
Some of the clustering platforms have options to handle outliers in the data. However, if your data has outliers, it is best to explore them first prior to analyzing. This can be done using the Explore Outliers Utility. For more information, see Explore Outliers in Predictive and Specialized Modeling.