References:
Lakkaraju et al. (2016), Lakkaraju et al. (2017), Lakkaraju et al. (2019), Moradi and Samwald (2021), Hosain et al. (2024)
Toggle Text Reference
The overlap within a rule-based explanans (e.g., rule sets or decision trees) can be measured with respect to either the input space or the rules themselves. A lower degree of overlap is generally preferred, as it implies more distinct, non-redundant rules and enhances interpretability.
• Input Overlap: Measures how often input instances are covered by multiple rules or decision paths [Lakkaraju et al. (2016), Lakkaraju et al. (2017), Lakkaraju et al. (2019), Hosain et al. (2024)]. High overlap may indicate redundant or conflicting logic. Overlap can also be broken down by class label to distinguish intra-class from inter-class coverage.
• Rule Overlap: Measures how many rules share identical or highly similar predicates [Moradi and Samwald (2021)], indicating structural redundancy within the rule set.
• Input Overlap: Measures how often input instances are covered by multiple rules or decision paths [Lakkaraju et al. (2016), Lakkaraju et al. (2017), Lakkaraju et al. (2019), Hosain et al. (2024)]. High overlap may indicate redundant or conflicting logic. Overlap can also be broken down by class label to distinguish intra-class from inter-class coverage.
• Rule Overlap: Measures how many rules share identical or highly similar predicates [Moradi and Samwald (2021)], indicating structural redundancy within the rule set.