Fast Communication-Efficient Spectral Clustering over Distributed Data

The last decades have seen a surge of interests in distributed computing thanks to advances clustered and big data technology. Existing algorithms typically assume {\it all the are already one place}, divide conquer on multiple machines. However, it is increasingly often that located at number sites, wishes compute over with low communication overhead. For spectral clustering, we propose novel framework enables its computation such data, "minimal" communications while major speedup computation. loss accuracy negligible compared non-distributed setting. Our approach allows local parallel where located, thus turns nature into blessing; most substantial when evenly across sites. Experiments synthetic large UC Irvine datasets show almost no our about 2x under various settings two As transmitted need not be their original form, readily addresses privacy concern for sharing computing.