Figure 10: Average difference in predictive accuracy between a model and its generated rulesets. The lower part with negative values is obtained when rules and network agree.