Learning with only the relevant features

In real-world concept-learning problems, the representation of data often uses many features, only a few of which may be related to the target concept. In this situation, the selection of relevant features becomes important for achieving high performance of a learning algorithm. A feature selection method based on tests has been introduced and its effectiveness in achieving high recognition rates has been shown (a test is a set of features that is sufficient to distinguish patterns from different classes of training samples). We show that the notion of a test is directly related to the notion of relevant attributes. We prove that prime tests represent sets of irredundant relevant features, and emphasize the importance of the test notion. We describe a way of selecting all relevant features from training examples, and possible ways to identify relevant features of the target concept. The empirical results show the effectiveness of the proposed method on real databases