This article presents several additive corrections to the conventional quadratic loss bias-plus-variance formula. One of these corrections is appropriate when both the target is not fixed (as in Bayesian analysis) and training sets are averaged over (as in the conventional bias plus variance formula). Another additive correction casts conventional fixed-trainingset Bayesian analysis directly in terms of bias plus variance. Another correction is appropriate for measuring full generalization error over a test set rather than (as with conventional bias plus variance) error at a single point. Yet another correction can help explain the recent counterintuitive bias-variance decomposition of Friedman for zero-one loss. After presenting these corrections, this article discusses some other loss function-specific aspects of supervised learning. In particular, there is a discussion of the fact that if the loss function is a metric (e.g., zero-one loss), then there is bound on the change in generalization error accompanying changing the algorithm's guess from h 1 to h 2 , a bound that depends only on h 1 and h 2 and not on the target. This article ends by presenting versions of the bias-plus-variance formula appropriate for logarithmic and quadratic scoring, and then all the additive corrections appropriate to those formulas. All the correction terms presented are a covariance, between the learning algorithm and the posterior distribution over targets. Accordingly, in the (very common) contexts in which those terms apply, there is not a “bias-variance trade-off” or a “bias-variance dilemma,” as one often hears. Rather there is a bias-variance-covariance trade-off.

This is the first of two papers that use off-training set (OTS) error to investigate the assumption-free relationship between learning algorithms. This first paper discusses the senses in which there are no a priori distinctions between learning algorithms. (The second paper discusses the senses in which there are such distinctions.) In this first paper it is shown, loosely speaking, that for any two algorithms A and B, there are “as many” targets (or priors over targets) for which A has lower expected OTS error than B as vice versa, for loss functions like zero-one loss. In particular, this is true if A is cross-validation and B is “anti-cross-validation” (choose the learning algorithm with largest cross-validation error). This paper ends with a discussion of the implications of these results for computational learning theory. It is shown that one cannot say: if empirical misclassification rate is low, the Vapnik-Chervonenkis dimension of your generalizer is small, and the training set is large, then with high probability your OTS error is small. Other implications for “membership queries” algorithms and “punting” algorithms are also discussed.

This is the second of two papers that use off-training set (OTS) error to investigate the assumption-free relationship between learning algorithms. The first paper discusses a particular set of ways to compare learning algorithms, according to which there are no distinctions between learning algorithms. This second paper concentrates on different ways of comparing learning algorithms from those used in the first paper. In particular this second paper discusses the associated a priori distinctions that do exist between learning algorithms. In this second paper it is shown, loosely speaking, that for loss functions other than zero-one (e.g., quadratic loss), there are a priori distinctions between algorithms. However, even for such loss functions, it is shown here that any algorithm is equivalent on average to its “randomized” version, and in this still has no first principles justification in terms of average error. Nonetheless, as this paper discusses, it may be that (for example) cross-validation has better head-to-head minimax properties than “anti-cross-validation” (choose the learning algorithm with the largest cross-validation error). This may be true even for zero-one loss, a loss function for which the notion of “randomization” would not be relevant. This paper also analyzes averages over hypotheses rather than targets. Such analyses hold for all possible priors over targets. Accordingly they prove, as a particular example, that cross-validation cannot be justified as a Bayesian procedure. In fact, for a very natural restriction of the class of learning algorithms, one should use anti-cross-validation rather than cross-validation (!).