No right answers here: find the structure in the data itself.
In many dimensions every point is about as far from every other.
With many dimensions "nearest" means little.
The direction with the most variance is the first principal component.
Where the bars drop to almost nothing, the rest is noise.
Assign every point to its nearest centre, move every centre to its points' mean, repeat.
Look for the elbow, and for the highest silhouette.