Based on our experimental setup, we tried to assess the predictive quality of the rules mined by AMIE and ALEPH, a state-of-the-art ILP learner. Since ALEPH could not run on the clean version of YAGO2, we ran both systems on a 47K sample with the minimum possible support (2 head bindings). ALEPH learns rules for one a relation at a time. Even on the sample, some relations did not terminate in more than one day and were therefore omitted from the experiment (for both systems). At the end we obtained 52 rules from ALEPH and 212* from AMIE which were ranked by their corresponding confidence scores (PCA confidence for AMIE, positive learning score for ALEPH). The aggregated predictions and precisions are plotted in the following chart:
As we can see, ALEPH produces good quality rules in terms of precision and recall. The top 15 rules for ALEPH produce as many predictions as the top 52 rules for AMIE. Note however, than AMIE achieves higher precision and also produces more rules. We still envision to go deeper into the insights of the ALEPH positive learning score function and compare it against our PCA confidence.
*335 rules in total