On the Trade-off Between Consistency and Coverage in Multi-label Rule Learning Heuristics