Abstract
The human brain forms a large-scale, interconnected network when performing different activities. To compare networks extracted from different subjects, they are first converted into sparse graphs with similar densities to reveal topological differences. Graph analysis is then applied to the sparse graphs to extract global and local graph invariants for quantitative comparisons. However, many previous works not only studied global and local graph invariants separately, but also created only one single sparse graph for each subject, potentially excluding important factors in connectome analysis. In this work, we adopt a more inclusive approach: generating multiple graphs using different density thresholds for each subject; and describing each graph with both global and local graph invariants. A machine learning approach is then applied to analyze these comprehensive datasets. We show that our inclusive approach can help machine learning methods to automatically identify most discriminating factors in predicting brain activities with much higher accuracy than the previous exclusive approaches.