0
$\begingroup$

I clustered some data (rows: text documents, columns: word frequencies) using the KMeans implementation in Scikit Learn. This, like most other centroid-based clustering algorithms, returns a centroid for each cluster.

I am now trying to identify features predictive of each cluster by simply comparing the value for a certain feature in a centroid to the other centroids. I just subtract the value of centroid 1's feature X from the value of centroid 2's feature X (cf. below) and then sort the differences to see which features have the widest spread and and are thus most strongly correlated with a cluster.

  • Does this approach make sense / work?
  • If not, why not and what should I do instead? (I found this but don't know how exhaustive it is)

I feel that the approach is kind of simplistic, but can't think of a much better way either.

More detailed description of idea, just for clarification:

In a 3 cluster k means, I have the vector for centroid 1 [1,2,0,...]. for centroid 2 [10, 2, 400,...] and centroid 3 [100, 2, 0, ...]. I then subtract feature 1: abs(1-10), abs(1-100), feature 2: abs(2-2), abs(2-2), etc. This I would interpret to mean that feature 1 is strongly predictive of cluster 3 membership, while feature 2 is not very predictive at all. In real life, a feature of course translates into a word. All my features are normalized as feature/total words. Any help is much appreciated!

$\endgroup$
1
$\begingroup$

If it helps, this is sort of how I built my talking teddy bear, which is an AI driven twitter engine of sorts (but its a physical bear).

The next step is of course, don't forget combinations of features being accelerators. There is a really interesting example on Microsoft's Azure Machine Learning gallery for this doing a clustering of similiar companies which provides a great starting point. Basically it does is feature hashes each companies wikipedia page into n-grams rolls up counts and fits them into various clusters. Works pretty darned well, I am using that strategy for identifing similiar companies for a financial application to diversify portfolios.

Advise: Start simple, if its not accurate enough, do more. If its accurate enough for the ROI, then stop, ship and start the next thing.

$\endgroup$
  • $\begingroup$ thanks! just so i understand correctly: you'd say the approach is valid, and to get even more accurate one could consider interactions/ correlations between features? $\endgroup$ – patrick Apr 28 '16 at 17:43
  • $\begingroup$ Exactly. I typically start with the simplest fastest approach and build from there if necessary. $\endgroup$ – David Crook Apr 28 '16 at 17:49
  • $\begingroup$ Just for good measure, check out the sample in Azure Machine Learnings gallery. Most of the sample is written in R, but is mostly stock R and therefor fairly easily transferrable to python. $\endgroup$ – David Crook Apr 28 '16 at 17:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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