A simple rule evaluates one condition from one feature and returns a binary fraud decision. It can use a binary feature directly or compare a numeric feature against a threshold. section 2.2.1 section 2.4
Use simple rules as an interpretable baseline before combining rules:
- Measure each rule against the fraud label with accuracy, while recognizing that baseline performance depends on class balance. section 2.2.1 section 2.4
- Treat feature direction carefully: one feature may indicate fraud at 1, while an inversely related feature indicates fraud at 0. section 2.2.1
- Combine simple rules with logical operators only after evaluating them individually; combinations become compound rules and can increase system complexity exponentially. section 2.2.1 section 2.4
Phishing URL baseline
Define rule_1 as FrequentDomainNameMismatch == 1. On the phishing dataset, this simple rule reaches 69.07% accuracy, compared with a 50% baseline for the balanced fraud and genuine classes. A second rule, SubmitInfoToEmail == 0, reaches 61.98%. section 2.2.1
rule_1 = FrequentDomainNameMismatch == 1
Inspect the minimal Python expression that turns FrequentDomainNameMismatch into a simple rule. section 2.2.1
