Purpose of dendrogram and hierarchical clustering

This is likely a very naive question. I've lately been reading about hierarchical clustering algorithms, and various discussions about how to interpret dendrograms or find optimal heights for cutting a dendrogram. I've also played around with some hierarchical clustering software in Python.

My question is: what exactly can a dendrogram tell you that you couldn't find out from a simple distance matrix?

Firstly, computing a dendrogram is very costly. Most linkage methods (such as group average linkage) have an algorithmic complexity of O(N^2) logN, which any comp-sci student can tell you, is pretty horrible in terms of scalability. Even a dataset with something like only 1,000 datapoints could be relatively slow on modern computing hardware.

Secondly, it seems that once you actually have the dendrogram, it's not always clear exactly how to cut it. However, generally you want to try and find clusters with lower distance values. Intuitively, clusters at heights with very high distance values are likely not very significant.

So with all that said, what exactly can a dendrogram tell you that you couldn't ascertain with a simple distance matrix? By that I mean, instead of constructing a dendrogram, simply compute a NxN distance matrix in O(N^2) time, then iterate over the matrix and find distance values which fall within some predefined range or threshold. This will give you a list of all points which are "close" to each other, and produce a single cluster of likely-related objects.

So, my question is, what can a dendrogram tell you that may potentially be more informative, significant or useful than a simple distance matrix?

• You've mixed two separate questions in one. Your 1st Q is actually about hierarchical cluster analysis, not dendrogram. (It is possible to do the HCA and save a range of cluster solutions and then choose the "best" one, - completely refraining from drawing the dendrogram.) Your second Q is about dendrogram, not HCA; moreover, that Q isn't quite clear, - can you be more specific: what bothers you wrt dendrogram. Here are some of my notes about HCA and dendrogram, for your interest. Sep 3, 2016 at 3:15
• In your last but one paragraph you seem to want to suggest some algorithm completely different from HCA, but you tell so little about its purposes and principles and about whether it can be a substitute for HCA - that it is difficult to comment on it. To sum up: you mix up three quite different topics in your question; maybe you should re-write your post to make it narrower and specific. Sep 3, 2016 at 3:20