We examine the mechanism by which feature selection improves the accuracy of supervised learning. An empirical bias/variance analysis as feature selection progresses indicates that the most accurate feature set corresponds to the best bias-variance trade-off point for the learning algorithm. Often, this is not the point separating relevant from irrelevant features, but where increasing variance outweighs the gains from adding more (weakly) relevant features. In other words, feature selection can be viewed as a variance reduction method that trades off the benefits of decreased variance (from the reduction in dimensionality) with the harm of increased bias (from eliminating some of the relevant features). If a variance reduction method like bagging is used, more (weakly) relevant features can be exploited and the most accurate feature set is usually larger. In many cases, the best performance is obtained by using all available features.