KNN clustering with categorical data

Thanks bogatron. How do I practically use this? At the moment, I'm implement KNN searching with a simple MATLAB function. The function calculates Euclidean distances to each neighbour in the cluster. So the algorithm would see, for e.g., male = 0; female = 1. However, I can't pass in three dummy variables for other categorical data. What to do?

I was thinking about it a bit more - are you saying that you create 3 NEW dummy variables which are purely binary now? Also, surely binary numbers have different weights (e.g., male = 0, female = 1), so this won't work with Euclidean distances? Is hamming a better option, and will it also work with numerical values?

If your numerical attributes are normalized, then you can map a binary attribute to (0,1). So you could have a gender attribute that takes the value of 0 or 1 when gender is male or female, respectively.

Yes, I was suggesting that you could replace "bank name" with three features (one for each value). The benefit of that approach is that it will easily fit into a standard clustering algorithm. The downside is that the dimensionality of your data can become very large if you have many bank names. You could also use hamming distance, as you suggested. That would keep the dimensionality of the data lower but you would have to explicitly code the knn algorithm to combine hamming distance with the "normal" numerical feature distances.