Figure 4

Comparison of learned dlink structure and true dlink structure. (A) This is a discretized version of Figure 2(A). For a given pair of matrices, we can characterize each edge as a true positive if it occurs in both the true and inferred dlink matrix, a true negative if it occurs in neither, or a false positive or false negative if it occurs only in the inferred or only in the true matrix, respectively. In the resulting sparse model, an edge (i.e., dlink) is counted as occurring when the absolute value of its dlink coefficient is greater than zero. Otherwise, it is counted as non-occurring. In the figure, each dlink is colored according to whether it is a true positive, false positive, true negative or false negative. Overall, this particular inference procedure yields 4 false positives, and 9 false negatives from a total of 2190 possible edges, for an accuracy of 2177/2190 = 99.4%. (B) The figure plots inference accuracy ((TP + FP)/(TP + FP + TN + FN)) as a function of training set size, for a fixed sparsity level of 20%. As the size of the training set grows, the algorithm is able to converge to the true sparse model.