Selection of relevant features and examples in machine learning

Reviewer: Daniel L. Chester

Machine learning problems are usually represented by examples of concepts; each example is a vector of feature values together with the classification to which this vector belongs. In many applications, particularly in data mining, the available data may provide too many features and examples for machine learning algorithms to handle. If the data can be reduced to only the relevant features and examples, however, the learning task can be tractable. This paper surveys advances in automating feature and example selection. After discussing algorithms that heuristically search for relevant features, the authors survey approaches that embed relevant feature search in machine learning algorithms, that apply the search as a filter to data before the data are passed to the learning algorithms, or that call learning algorithms as subroutines within the search for relevant features. They also discuss weighting schemes that reflect how relevant each feature might be to the concept being learned. In a shorter discussion, they survey approaches to reducing the amount of processing that is done on examples. Many of these approaches focus on the examples that probe near the borders of the concept being learned. Anyone who wants to apply machine learning to large amounts of data will want to read this paper and look up the references for details on the algorithms that are mentioned. Researchers will be interested in the challenging problems offered at the end of the paper.