Java: Data Science Made Easy

Clusters as features

Clustering can be seen as a method for feature engineering, and the results of clustering can be added to a supervised model as a set of additional features.

The simplest way of doing it to use one-hot-encoding of clustering results is as follows:

• First, you run a clustering algorithm and as a result, you group the dataset into K clusters
• Then, you represent each datapoint as a cluster to which it belongs using the cluster ID
• Finally, you treat the IDs as a categorical feature and apply One-Hot-Encoding to it.

It looks very simple in code:

KMeans km = new KMeans(X, k, maxIter, runs); int[] labels = km.getClusterLabel(); SparseDataset sparse = new SparseDataset(k); for (int i = 0; i < labels.length; i++) {  sparse.set(i, ...