It appears that when data sets have a combination of categorical and continuous attributes, the common way to apply kernel algorithms to such data sets is to use a one hot encoding scheme for each categorical variable - this requires replacing a categorical variable with K levels with K dummy indicator variables. I am interested in unsupervised applications of kernel methods. I tried the above idea (performing PCA) on a data set with mixed attributes using an RBF kernel but the eigenvalues I got did not have the desirable scree plot look. I am aware of Multiple Correspondence Analysis, I am just curious if there is a way to do this using kernel methods - perhaps a better encoding scheme or better choice of kernel etc. I'd appreciate any insights in solving this problem with kernels.
I'm not an expert on mixed data but there are indeed some kernels that combine continuous and categorical attributes, such as the clinical kernel. In the paper where the clinical kernel was proposed you will most likely find other approaches and proposals in the references. There are also some follow up papers using that kernel with good results.