Dear Gunjan,

It is really a problem clustering large data. Of course parallel computing can solve the issue by speeding up the process of clustering. However, the same steps that should be performed in single processor are to be performed in multi processors. The difference is the speed of the process with multi processors. In general you have to subdivide the data into sub matrices of 6400X6400 this means you have 9 sub matrices. Then do the clustering for each separately, but at the end you will have the problem of recombining them and this problem can be solved easily.

Best regards

.

the basic building block of spectral clustering is a svd

before going "parallel" (maybe you only mean "distributed", hadoop style) you might be interested in trying approximate svd techniques (spectral clustering does not require the accurate computation of all singular values) ; see the attached papers by Liberty and their bibliography (note that hash-based techniques are not as bad as they say because you can use larger sketches at virtually no cost as long as you have enough samples to "fill" them and that the sketch matrix size allows the final step svd computation)

.

otherwise, if you want to go "distributed", the following reference might be useful

http://web.cs.ucla.edu/~ameet/dfc_jmlr.pdf

.

Dear Gunjan,

Please check more information about this issue in the attached paper.

Thanks M.Awad and Fabrice Clerot for your valuable opinion.I will look into this and let you know what I am getting.