The paper introduces the concept of instance-based learning (IBL) algorithms, which use specific instances to generate classification predictions without maintaining abstractions. IBL algorithms are derived from the nearest neighbor pattern classifier and are similar to edited nearest neighbor algorithms, which also save and use selected instances. The authors describe a framework for IBL algorithms, including a similarity function, a classification function, and a concept description updater. They analyze the generality, accuracy, learning rate, incorporation costs, and storage requirements of IBL algorithms. The simplest IBL algorithm, IBI, is presented, and its performance is compared with other algorithms. The paper also discusses the IB2 algorithm, which reduces storage requirements by saving only misclassified instances, and the IB3 algorithm, which tolerates noisy instances using a significance test. Empirical studies show that IB2 significantly reduces storage requirements but may sacrifice classification accuracy, especially in noisy or complex domains. IB3 addresses this issue by selectively using instances with high classification accuracy. The authors conclude by discussing the limitations and advantages of IBL algorithms and highlighting areas for further research.The paper introduces the concept of instance-based learning (IBL) algorithms, which use specific instances to generate classification predictions without maintaining abstractions. IBL algorithms are derived from the nearest neighbor pattern classifier and are similar to edited nearest neighbor algorithms, which also save and use selected instances. The authors describe a framework for IBL algorithms, including a similarity function, a classification function, and a concept description updater. They analyze the generality, accuracy, learning rate, incorporation costs, and storage requirements of IBL algorithms. The simplest IBL algorithm, IBI, is presented, and its performance is compared with other algorithms. The paper also discusses the IB2 algorithm, which reduces storage requirements by saving only misclassified instances, and the IB3 algorithm, which tolerates noisy instances using a significance test. Empirical studies show that IB2 significantly reduces storage requirements but may sacrifice classification accuracy, especially in noisy or complex domains. IB3 addresses this issue by selectively using instances with high classification accuracy. The authors conclude by discussing the limitations and advantages of IBL algorithms and highlighting areas for further research.