Feature weighting is known empirically to improve classification accuracy for k-nearest neighbor classifiers in tasks with irrelevant features. Many feature weighting algorithms are designed to work with symbolic features, or numeric features, or both, but cannot be applied to problems with features that do not fit these categories. This paper presents a new k-nearest neighbor feature weighting algorithm that works with any kind of feature for which a distance function can be defined. Applied to an image classification task with unusual set-like features, the technique improves classification accuracy significantly. In tests on standard data sets from the UCI repository, the technique yields improvements comparable to weighting features by information gain.