Given two sets of nucleotide n-mers, how can I learn sets of 'motifs' which are present in one and not the other?

Question

[mk]

I have some unpublished data on DNA methylation, suggesting that there are differences in maintenance methylation efficiency at different sites from one cell cycle to another. I have a very coarse-grained approach for dividing the methylated sites into "likes to be methylated" and "really likes to be methylated". What is the best approach (perhaps some commoditized machine learning available in R) to generate sets of possible 'preferred motifs' given the sequences surrounding the fast and slow sites?