Tell me more ×
Cross Validated is a question and answer site for statisticians, data analysts, data miners and data visualization experts. It's 100% free, no registration required.
up vote 2 down vote favorite
1

In supervised learning, there are some ensemble methods that overcome others significantly (adaboost or random forests to mention some).

Few years later, also ensembles in unsupervised learning were proposed. Are there any techniques that outperform the others? What are your experiences with them w.r.t. their accuracy, easy manipulation and time complexity?

share|improve this question
1  
Just a general comment on unsupervised learning: In general it is very difficult to compare methods in this field, since you have to define what you mean by a "good clustering". This is a problem because not only can you define this in multiple ways, but this often implies some assumption on the distribution that generated the data (e.g. Gaussian), which is generally a bad thing. What most people do is compare either by using one of these definitions (the one that gives the best results for their algorithm?) or compare on data that have some external labels - both of these methods are flawed. – Bitwise Oct 9 '12 at 16:09
I agree with you, so we could specify good clustering by for example any external information (widely used is Iris dataset, where true labels are known) or by any internal validity criterion (average silhouette for example). – Miroslav Sabo Oct 9 '12 at 16:14
my point is, that I bet that any consensus clustering paper that is published shows it is better than other methods under some dataset or some clustering performance metric. So your question about which ones outperform the others is difficult to answer. I would try looking for some "unbiased" review paper on this type of methods. – Bitwise Oct 9 '12 at 16:22
1  
Supervised learning is very different because there you have a mathematically well-defined goal, which is to predict the labels as well as possible. That makes comparison relatively straightforward and unbiased (this is also why I somewhat dislike unsupervised learning). In clustering it is unnatural to use labels, because if you had labels you would probably be using supervised learning... – Bitwise Oct 9 '12 at 16:44
1  
Internal evaluation is not the best "unsupervised" approach. Essentially, you are making it a "self-supervised" approach, supervised by the criterion. k-means is a good example of this. It does a reasonably good job at optimizing a particular criterion. But it's also very easy to construct a data set, where the "optimal" result is completely useless. You do want to a large extend to have the evaluation driven by "usefulness", which unfortunately needs some kind of external evaluation. Yet, with labels, we could as well do supervised learning. So we need yet something else. – Anony-Mousse Oct 9 '12 at 20:04
show 2 more comments

1 Answer

up vote 1 down vote

Ensembles in clustering are not that well researched yet.

In supervised learning, you can use your target knowledge to optimize ensembles. However, how do you do unsupervised optimization?

There are some approaches on trying to maximize diversity while also looking for consensus. But the problem here is that a lot of the methods are strict, and their results are not really comparable. So a consensus may either be something too obvious or even an indication that both methods failed, at which point you didn't gain anything, actually.

The problem is here, that clustering is not just (unsupervised) learning. Actually it is a lot of "unsupervised" and not really learning. You need to look at it from a "knowledge discovery" (tell me something new about my data) view, not a "machine learning" viewpoint (label new data just like we already did with the previous data). It's really a different mind set.

Sorry, there is no "best" solution, because IMHO there is no "good" solution yet.

There is quite some research going on though. I have not been following all the MultiClust workshops, but these are highly relevant for what you are looking for.

share|improve this answer
Thank you, your answer is helpfull. – Miroslav Sabo Oct 9 '12 at 17:12

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.