Online and batch learning of pseudo-metrics

摘要: We describe and analyze an online algorithm for supervised learning of pseudo-metrics. The receives pairs instances predicts their similarity according to a pseudo-metric. pseudo-metrics we use are quadratic forms parameterized by positive semi-definite matrices. core the is update rule that based on successive projections onto cone half-space constraints imposed examples. efficient procedure performing these projections, derive worst case mistake bound predictions, discuss dual version in which it simple incorporate kernel operators. also serves as building block deriving large-margin batch algorithm. demonstrate merits proposed approach conducting experiments MNIST dataset document filtering.