Abstract
In clustering and other unsupervised tasks, one often seeks information about the data from the top eigenvectors of a graph-based operator. However, these may not always be the informative eigenvectors, due to various outliers, resulting from degree variations, tangles and more. Graph powering is a technique that tries to modify the graph to get rid of such outliers and bring back the informative eigenvectors at the top. We will argue that powering can handle both stochastic block models, and some 'geometric block model' where short loops are much more present. Joint work with C. Sandon and E. Boix.