We are trying a DBSCAN clustering model on our 30,000 samples with 15 features each. We tuned the epsilon parameter small enough to make sure the radius of the clustering circle is small while it does clustering. We expected to have small clusters with high density. After clustering and parameters tuning, we used t-SNE to plot the clustering results in 2 dimensional space, we found that we have small clusters like cluster 2,3,4,5 with high density as expected while large clusters like cluster 0,1 scattered loosely as unexpected. obviously, cluster 0, 1 looks very strange to us.
Is there something we are not aware of, is it something wrong with clustering algorithm?
Or it's something about t-SNE, when it plots the clustering results in lower dimension, it might look like cluster 0, 1 are very scattered, actually they are more close and tensed in higher dimensional space?
- cluster 0: orange
- cluster 1: blue
- cluster 2: red
- cluster 3: green
- cluster 4: light yellow
- cluster 5: purple
30,000 samples, 6 clusters case:
30,000 samples, 8 clusters case: