Оценивание сходства разбиений графов на пересекающиеся сообщества
In this work, we analyze the structure of connectome communities - graphs built on the basis of neuroimaging data of the human brain. We consider three approaches to the definition of cluster structures on connectomes: non-overlapping, overlapping, and fuzzy communities. For each of the approaches, we analyze the corresponding metric for assessing the similarity of partitions on connectomes: the Rand index and its generalizations for overlapping communities - the Omega index and the fuzzy Rand index, differing in the assumption of the degree of belonging of a vertex to several communities at the same time. We conduct a comparative analysis of these approaches, evaluating them from the point of view of information content for subsequent classification, using the example of patients with and without Alzheimer's disease. We use k-nearest neighbors and support vector machines as classifiers. The best results in the classification problem were obtained when considering intersecting communities using a support vector machine (ROC AUC 0.83).