Fig. 4

The discriminative model usually improves faster than the generative model as more edge-specific label functions are included. The line plot headers represent the specific edge type the discriminative model is trying to predict. The x-axis shows the number of randomly sampled label functions incorporated as an addition to the baseline model (the point at 0). The y axis shows the area under the receiver operating curve (AUROC). Each data point represents the average of 3 sample runs for the discriminator model and 50 sample runs for the generative model. The error bars represent each run’s 95% confidence interval. The baseline and “All” data points consist of sampling from the entire fixed set of label functions