关键词: 多视角学习, 自适应, L1-图, 图像分类

Abstract: Since the existing multi-view learning algorithms often suffer from the problem of lacking data adaptiveness in the nearest neighbor graph construction procedure, an Adaptive Multi-View Learning (AMVL) algorithm was proposed. Firstly, by utilizing the automatic data sample selection property of L1 norm constraint, multiple view-related directed L1-graphs were constructed. Secondly, according to the obtained L1 graphs, the algorithm tried to minimize the low dimensional reconstruction error in each view. Lastly, the objective function of the proposed adaptive multi-view learning algorithm was obtained by performing global coordinate alignment process in different views. Moreover, an iterative optimization method was also proposed to solve the proposed objective function. The algorithm was applied to the problem of image classification on two public image datasets, i.e., Corel5K and NUS-WIDE-OBJECT, and compared with several existing methods. The experimental results show that： a) the proposed algorithm can increase the classification accuracy up to 5% and 2% respectively on these two datasets; b) the optimization method convergences within 100 iterations; c) the number of nearest neighborhoods learned by the algorithm is datum-adaptive.

Key words: multi-view learning, adaptiveness, L1-graph, image classification