Exceptional Attributed Subgraph Mining to Understand the Olfactory Percept

Human olfactory perception is a complex phenomenon whose neural mechanisms are still largely unknown and novel methods are needed to better understand it. Methodological issues that prevent such understanding are: (1) to be comparable, individual cerebral images have to be transformed in order to fit a template brain, leading to a spatial imprecision that has to be taken into account in the analysis; (2) we have to deal with inter-individual variability of the hemodynamic signal from fMRI images which render comparisons of individual raw data difficult. The aim of the present paper was to overcome these issues. To this end, we developed a methodology based on discovering exceptional attributed subgraphs which enabled extracting invariants from fMRI data of a sample of individuals breathing different odorant molecules.Four attributed graph models were proposed that differ in how they report the hemodynamic activity measured in each voxel by associating varied attributes to the vertices of the graph. An extensive empirical study is presented that compares the ability of each modeling to uncover some brain areas that are of interest for the neuroscientists.