I have a dataset with categorical features. I want to segment the data using clustering techniques. What could be the possible choices for this scenariogiven the fact that data has categorical features. Is there any variation of k-means which can be used here.
$\begingroup$ If your categorical features exhibit an order, you transform these features by assigning a number to each level. If it is not the case, you could add one feature per value, and assign a binary value to it. This way, those feature would be orthogonal. $\endgroup$– Manu HAug 8, 2016 at 9:15
$\begingroup$ The categorical Feature has more than 1000 different categories and does not exhibit any order. I guess the above Approach would then make the task computationally expensive. $\endgroup$– user3198880Aug 8, 2016 at 12:29
$\begingroup$ it will indeed be memory expensive. I would try a kernel-trick ( seems to work with Kmeans ) with an approahc similar to the MKL, the value of the kernel for the categorical part being zero if the category is not the same, one otherwise. not sure my thought is clear. $\endgroup$– Manu HAug 8, 2016 at 15:50
k-means is not a good choice, because it is designed for continuous variables. It is a least-squares problem definition - a deviation of 2.0 is 4x as bad as a deviation of 1.0.
On binary data (such as one-hot encoded categorical data), this notion of squared deviations is not very appropriate. In particular, the cluster centroids are not binary vectors anymore!
The question you should ask first is: "what is a cluster". Don't just hope an algorithm works. Choose (or build!) and algorithm that solves your problem, not someone else's!
On categorical data, frequent itemsets are usually the much better concept of a cluster than the centroid concept of k-means.
Not enough reputation to comment...
Do you have any insight on whether your categorical variables exhibit some ordering? Or are they nominal? Is it possible to impose an ordering on your variables such that it is intuitive?
Your problem comes down to choosing an appropriate distance metric. Or rather, what defines 'similarity' to you. There is a variant of the k-means algorithm called k-modes that you may want to explore. The last link below provides more information on this categorical clustering method.
In absence of knowing more about your data, these links might be useful:
I don't really see a reason why simple K-Means clustering shouldn't work. If you convert your categorical data into integers (or encode to binary where one column is equal to one category, so called "one-hot encoding"), you can then fetch it into the algorithm.
Then, you can compare the cluster between each other by, lets say, calculate the mode to see the differences.
Also, as dmanuge mentioned, playing with different metric can be helpful. But I'd go for this after the simple K-Means.
$\begingroup$ I have one column in the data set which is categorical and has more than 1000 different Levels or categories. $\endgroup$ Aug 8, 2016 at 9:09
$\begingroup$ Then I'd first make sure you really need 1000 categories (there is really business process with 1000 categories?), then I'd merge some together. Also validate if the column is useful at all ("feature selection"). After that, you can do one-hot encoding. Or if you aim only at certain category, try "One-vs-All" approach. $\endgroup$– HonzaBAug 8, 2016 at 9:40
Your approach may depend on the number of features and the number of categories in each feature that you are trying to include in your model. I've used dummy variables to convert categorical data into numerical data and then used the dummy variables to do K-means clustering with some success.
Here's a small example:
+----+----+----+ | ID | F1 | F2 | +----+----+----+ | 1 | a | x | | 2 | d | w | | 3 | f | x | +----+----+----+
Create a column for each category of each feature. For each record, the value of the dummy variable field is 1 only in the dummy variable field that corresponds to the initial feature value. The rest are 0.
+----+------+------+------+------+------+ | ID | F1_a | F1_d | F1_f | F2_w | F2_x | +----+------+------+------+------+------+ | 1 | 1 | 0 | 0 | 0 | 1 | | 2 | 0 | 1 | 0 | 1 | 0 | | 3 | 0 | 0 | 1 | 0 | 1 | +----+------+------+------+------+------+
If you're working with Pandas in Python,
pandas.get_dummies() can generate the dummy variables for you.
Sometimes, you could have so many categories it would been unreasonable to try and create a dummy variable for each one. For my problem, it was acceptable to only include in my model dummy variables for categories that occurred most frequently (e.g. Top 15 categories), but you'll have to decide whether or not that's appropriate for your problem.