Qualitative and quantitative description of functional connectivity graphs using graph attributes is of great interest to neuroscience, and has led to remarkable insights in the field. However, the statistical techniques used have generally been limited to whole-group, post-hoc studies. In this paper, we propose instead a novel approach to perform predictive inference on single subjects. It is based on a lossy embedding of connectivity graphs into a vector space using graph and vertex attributes, followed by the use of statistical machine learning to build a predictive model. The feature space proposed is easily interpretable for neuroscientists, and we illustrate the technique by revealing resting-state difference between young and elderly subjects.