Without reading the article at all, this kind of thing sounds like an interesting mathematical problem. You have some time series above various nodes on a grap

Without reading the article at all, this kind of thing sounds like an interesting mathematical problem. You have some time series above various nodes on a graph (each node is an account), and you cannot observe the time series directly, only some functions of the total state space, such as yearly income statements, ( i.e. some information is hidden or obfuscated). The question is how much of the underlying graph and time series you can reconstruct from your observations. I know some people have methods for reconstructing time dependent dynamical systems from observations using E.g. manifold learning, but I'm not sure if this problem exactly fits the bill.