Figure 2
From: Learning sparse models for a dynamic Bayesian network classifier of protein secondary structure
Comparison of learned and true dlink values. (A) The figure depicts the difference between the dlink values learned from real data and synthetic data with 30,000 proteins for helices at 20% sparsity level (see Figure 4(A) for a more detailed representation of this result). (B) The figure plots, as a function of training set size, the mean of the MAD metric (see text) computed across ten replicate experiments with 20% sparsity. The three series correspond to the helix, strand, and loop graphs. Error bars correspond to standard deviations. Both figures demonstrate that the model parameters learned from synthetic data are close to the parameters learned from real data. This shows that when the data is generated from a sparse model (i.e., the true model we are trying to recover is sparse) the sparsity algorithm proposed in this paper is able to learn these parameters correctly.