Plots depicting the results of a k-means clustering algorithm with four clusters using the best feature set. On the left is a BPT diagram, and on the right is a the two-dimensional embedding of our data as determined by t-SNE. Although we use four clusters, one of the clusters is empty, so this is effectively k = 3, but still achieves better accuracy than any model we trained with k = 3. The overwhelming majority of the sample fall into a single cluster, which comprises nearly the entire inactive region of the t-SNE diagram, but also the majority of the active region (left of linear cut).