@Ben: there's an example included in readme.txt
it's simple, you only need to compute distance_matrix by yourself. E.g., you can download MNIST data from the link in my paper, and compute Euclidean distance matrix. distance_matrix is n x n if there're n samples. groupNumber is number of clusters. K is for K-NN graph.

In the implementation, should we feed a dense graphW matrix other than a sparse matrix? Since there is a lot of zeros if the graph is constructed by the knn strategy, full(graphW) will waste a lot of storage...

Thanks for the code. BTW, I find that in your algorithm, the group number must be determined, but not adaptive, like meanshift, right? Could you please give me some suggestions to improve your algorithm if I can not provide a determined group number? Thank you very much.

In the implementation, should we feed a dense graphW matrix other than a sparse matrix? Since there is a lot of zeros if the graph is constructed by the knn strategy, full(graphW) will waste a lot of storage...

Thanks for the code. BTW, I find that in your algorithm, the group number must be determined, but not adaptive, like meanshift, right? Could you please give me some suggestions to improve your algorithm if I can not provide a determined group number? Thank you very much.

Comment only