Fast LMNN Algorithm through Random Sampling

The Large Margin Nearest Neighbor (LMNN) metric learning algorithm has been successfully used in many applications and continuously motivates new research works. However, the high computational complexity of training the LMNN algorithm inherent from the k-Nearest Neighbor (kNN) search makes it inapplicable to large datasets, especially when we need to tune the hyper-parameters of the LMNN algorithm. In this paper, we present the fast LMNN algorithm through random sampling. Random sampling method reduces the number of samples that needs to be considered and therefore greatly reduces the computational complexity of training the LMNN algorithm. Our experiments show that when the sample rate is 10%, the performance of LMNN algorithm is nearly the same to training on all data samples while the training time is only 8% to 40% of training on all data samples. We further show that random sampling method can efficiently tune the hyper-parameters of the LMNN algorithm.