Cool new paper: (Messe et al., 2015) "Predicting functional connectivity from structural connectivity..."

[whock]

An interesting paper recently came out in Neuroimage (link below). The authors review several computational models using graph theory to predict functional connectivity from structural connectivity.

http://www.sciencedirect.com/science/article/pii/S1053811915000932

1. The simplest model (SAR - simultaneous autoregressive model) did best. Is that surprising? Does that maybe have something to say about overfitting?
2. Also what do people think about what parameters the different models use to predict FC versus what you might think are good parameters?
3. The authors use a graph theoretic approach to compare empirical and simulated functional connectivities. Each connectivity result was treated as a node on a graph and (if I understand correctly) the authors used different metrics to see how similar predicted and empirical FCs were via weighted edges. They then analyzed the graphs using different graph theoretic properties to see how close the predicted/empirical FCs were under different models (I think). What do people think of this method? Is it sound? Would you have used different graph theoretic metrics?

[dlee138]

Regarding question 1, I was somewhat surprised to see the simplest model being the best. Although the results show that complex models are susceptible to overfitting, I thought the simplest model would be susceptible to underfitting. Therefore, I expected that a "not too simple, but not too complex" model would work the best, which wasn't the case in this paper. SAR was also a purely spatial with no dynamics, so I guess fewer parameters are better in some cases?

[mblohr]

Having not read the paper, my guess is that the simpler model performed best due to a small training data set.