This paper presents a new rule-induction method called PART, which learns accurate and compact rule sets without requiring global optimization. Unlike C4.5 and RIPPER, which use global optimization to refine initial rule sets, PART combines two rule-learning paradigms: creating rules from decision trees and the separate-and-conquer approach. It builds partial decision trees, then extracts a single rule from each. This approach avoids the need for global optimization, leading to simpler and more efficient rule sets. PART works by recursively building partial decision trees. Each tree is expanded until a subset of examples is split into leaves, which are then used to generate rules. The algorithm prunes the tree only when all children are leaves, preventing over-pruning. This ensures that rules are generalized only after their implications are known, avoiding hasty generalization. PART is efficient and produces rule sets that are as accurate as C4.5 and more accurate than RIPPER. It also performs well on datasets with noise and numeric attributes. Experiments on standard datasets show that PART outperforms C4.5 on nine datasets, is less accurate than C5.0 on ten datasets, and outperforms RIPPER on fourteen datasets. The size of the rule sets produced by PART is also smaller than those of C4.5 and C5.0 on many datasets. The algorithm handles missing attribute values by assigning weights to instances based on the number of training instances going down each branch. It also efficiently combines the weights of multiple rules for classification. The runtime of PART is efficient, with subquadratic performance even on large datasets. The paper concludes that PART is a simple and effective method for learning decision lists, combining two rule-learning paradigms to produce accurate and compact rule sets without the need for global optimization. Future research could explore whether using stopping criteria based on the minimum description length principle or reduced error pruning could further improve the size of the rule sets produced by PART.This paper presents a new rule-induction method called PART, which learns accurate and compact rule sets without requiring global optimization. Unlike C4.5 and RIPPER, which use global optimization to refine initial rule sets, PART combines two rule-learning paradigms: creating rules from decision trees and the separate-and-conquer approach. It builds partial decision trees, then extracts a single rule from each. This approach avoids the need for global optimization, leading to simpler and more efficient rule sets. PART works by recursively building partial decision trees. Each tree is expanded until a subset of examples is split into leaves, which are then used to generate rules. The algorithm prunes the tree only when all children are leaves, preventing over-pruning. This ensures that rules are generalized only after their implications are known, avoiding hasty generalization. PART is efficient and produces rule sets that are as accurate as C4.5 and more accurate than RIPPER. It also performs well on datasets with noise and numeric attributes. Experiments on standard datasets show that PART outperforms C4.5 on nine datasets, is less accurate than C5.0 on ten datasets, and outperforms RIPPER on fourteen datasets. The size of the rule sets produced by PART is also smaller than those of C4.5 and C5.0 on many datasets. The algorithm handles missing attribute values by assigning weights to instances based on the number of training instances going down each branch. It also efficiently combines the weights of multiple rules for classification. The runtime of PART is efficient, with subquadratic performance even on large datasets. The paper concludes that PART is a simple and effective method for learning decision lists, combining two rule-learning paradigms to produce accurate and compact rule sets without the need for global optimization. Future research could explore whether using stopping criteria based on the minimum description length principle or reduced error pruning could further improve the size of the rule sets produced by PART.