Regularized online learning of pseudometrics

We present a regularized approach for online learning of a pseudometric in the form of a Mahalanobis distance. We express the problem as an optimization that learns on the current labeled instance whilst favoring a solution of a predefined form. Our focus is on regularization. Our formulation takes up a flexible form allowing for scenarios ranging from traditional L2 regularization to regularization to a prior estimated from unsupervised data. We apply our method to an online content-based music retrieval scenario (e.g. personalized internet radio). Here the user provides information on his listening preferences via online feedback for each song that is played. By updating a pseudometric given this feedback, the algorithm optimizes a transformation that maps the user´s preferred songs closer together and undesired songs far from these preferred songs.