A Novel Efficient Discriminative Projection Learning for Big Data

For big data, hashing methods are currently utilized to learn hash codes; however, the performance of hashing methods degenerates rapidly as the code length increases. In contrast to the existing hashing methods, we develop a novel feature learning method named discriminant projection for big data. Specifically, the proposed projection learning method consists of two processes: optimal subset learning and discriminative projection learning. Optimal subset learning is applied to obtain a small-scale dataset; this reduces the amount of training dataset s and improves training efficiency. Discriminative projection learning is applied to learn the discriminant projection; this can make the same class samples more compact, and keep different class samples far away from each other. The proposed framework considers the discriminant information of the training data. Results of the experimental conducted with four databases clearly demonstrate the superior performance of the proposed framework.