Induction of Decision Trees from Inconclusive Data

Abstract Learning from inconclusive data is an important problem that has not been addressed in the concept learning literature. In this paper, we define inconclusiveness and illustrate why ID3-like algorithms are bound to result in overspecialized classifications when trained on inconclusive data. We address the difficult problem of deciding when to stop specialization during top-down decision tree generation, and describe a modified version of Quinlan's ID3 algorithm, called INFERULE, which addresses some of the problems involved in learning from inconclusive data. Results show that INFERULE outperformed ID3 (with and without pruning) in tests on a real-world diagnostic database containing automobile repair cases.