Boosting connectome classification via combination of geometric and topological normalizations

Simultaneous Matrix Diagonalization for Structural Brain Networks Classification

“…The possible explanation is in much higher dimensionality combined with greater noise in the data. The latter statement is partially supported by the fact that typical quality of classification for both datasets [21,9] is much lower than for ADNI2 database considered in this work.…”

“…The problem of brain network classification has been paid much attention recently [17,9,21,19,18]. This problem is non-trivial as most modern classification algorithms can work only with vectorial data while in our case each object in the dataset is represented by graph.…”

“…The first important question is what representation of a graph to use for further processing. The paper [21] considers different types of matrices such as adjacency matrix, graph Laplacian, normalized graph Laplacian and others. They further consider different types of processing for the chosen matrix:…”

“…Each of the choices above has certain pros and cons, however the authors of [21] report that the best results are achieved by combination of geometric and weighted communicability normalization.…”

“…The second important question is what to do with the matrix in order to use it as an input to a classification algorithm. Several approaches have been considered in the literature, including vectorization of the whole matrix, considering the vector of node degrees of a graph [21] or the vector of eigenvalues [9], and different variants of constructing the kernel between graphs [17,9]. We note, that in all these approaches features are generated for each connectome independently.…”