In principle one could penalize the likelihood in the EM algorithm by whatever seems to be reasonable considering the target application, but my understanding is that it is not anymore guaranteed that the EM algorithm will converge. This will of course depend on what exact kind of penalization one uses and how strong the contribution of the penalty term has. An example I have in mind is one where the likelihood is penalized by a Markov random field type term that forces EM to preferentially group neighboring pixels together. This type of penalization leads to problems with convegence if the penalty term is too large (i.e. the Gibbs field strength is too large).

Hrm. This paper is somwhat difficult to read, as situation which is somewhat exasperated by the fact that part of the PDF blinks. Does anyone else have this problem? (Perhaps I sholud sleep some and try again when it's light outside)

Question:
1.What is the disadvantage of the nonsymmetric dissimilarity measurement?

2.Will the feature space selected in the dissimilarity measurement influence the performance? If not, it is really good, but if yes, I think when using such measurements in real application, such as Anand's idea about combining it with the EM algorithm, the key becomes the feature selection part, which is also an important issue in clustering.
 
I think one of the most contribution of the paper is its systematic methodology in dealing with benchmark comparison stuff.

I wish to know if the dissimilarity measures mentioned in the paper can be used to improve the performance of existing clustering algorithms like the EM algorithm by regularization. It is well known that the EM tries to maximize the likelihood function. We can modify the likelihood to obtain a penalized likelihood where the penalizing term is a function of the dissimilarity measure between clusters. By maximizing this modified objective function, we can obtain clustering algorithms that provide better segmentation.

Any thoughts on this ?