I have a populations of 500 elements. Each element is represented by a 10 dimension feature vector which sum of element is equal to 1 (you can think about it as a histogram of probabilities). In particular the elements of this feature vector follow a Dirichelet distribution. ...OK they are the topics extracted by latent dirichelet allocation (LDA) :)

Is there any classifier that is particularly indicated for this kind of features?



If you want to build up the clustering model, refer to the Dirichlet process (DP). You need first go through the Dirichlet distribution and the posterior Dirichlet distribution(multinomial is the likelihood with the Dirichlet prior), then you will get a better understanding on the DP.

In terms of your question, if you think each of the data only belongs to one cluster, this model can be handled by Chinese Restaurant process (CRP). This is a model induced from the DP. You can think of the CRP as a large matrix, each row is a data point (500 elements in your case), each column is a cluster. In the theoretical CRP, the column number is countable infinite, this means that the cluster can grow when the sample size increases. Each of the data point can only be assigned to one cluster. You have 10 features, I think those 10 features can determine which cluster the data belongs to. Moreover, if each data point belongs to multiple features, you can use Indian Buffet Process.


Or you can refer to the last chapter of the book Bayesian Data Analysis, but I think the link above is more straight forward. http://www.stat.columbia.edu/~gelman/book/


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.