Beyond the arithmetic mean: extensions of spectral clustering and semi-supervised learning for signed and multilayer graphs via matrix power means

Pedro Eduardo Mercado Lopez

Abstract: In this thesis we present extensions of spectral clustering and semi-supervised learning to signed and multilayer graphs. These extensions are based on a one-parameter family of matrix functions called Matrix Power Means. In the scalar case, this family has the arithmetic, geometric and harmonic means as particular cases. We study the effectivity of this family of matrix functions through suitable versions of the stochastic block model to signed and multilayer graphs. We provide provable properties in expectation and further identify regimes where the state of the art fails whereas our approach provably performs well. Some of the settings that we analyze are as follows: first, the case where each layer presents a reliable approximation to the overall clustering; second, the case when one single layer has information about the clusters whereas the remaining layers are potentially just noise; third, the case when each layer has only partial information but all together show global information about the underlying clustering structure. We present extensive numerical verifications of all our results and provide matrix-free numerical schemes. With these numerical schemes we are able to show that our proposed approach based on matrix power means is scalable to large sparse signed and multilayer graphs. Finally, we evaluate our methods in real world datasets. For instance, we show that our approach consistently identifies clustering structure in a real signed network where previous approaches failed. This further verifies that our methods are competitive to the state of the art.