Graph-based Semi-supervised Dictionary Learning

In recent years,supervised dictionary learning has gained wide attention from researchers because of its better classification effect.The main idea is to extract the discriminative characteristics between classes by learning the training samples with labels,so as to provide a better basis for classification.However,it is very difficult to label a large number of samples,and it is very easy to collect a large number of unlabeled samples.Therefore,the rational use of the intrinsic structural relationship between the labeled and unlabeled samples is the key to improving the ability and ability to distinguish the dictionary.Based on this idea,two kinds of semi supervised dictionary learning methods based on graph structure are proposed in this paper.The main research work and contribution of this article can be summed up as follows:(1)A learning method of Graph-Based L1 Norm Semi-SupervisedDictionary Learning is proposed,the specific idea is:using sparse encoding between training samples and special label propagation method(Special Label Propagation,SLP)graph structure between training samples to construct sample obtained by soft label,and embedded into the dictionary learning framework,through graph structure constraints and sparse similar samples,forcing the unlabeled samples in the process of learning in the dictionary can be automatically added to the sample categories,and with the same kind of labeled samples share few dictionary atoms,so as to improve the dictionary sparse expression and discriminant ability.The experimental results on four data sets show that the algorithm has a good classification effect.(2)A large number of studies have shown that the l_p norm has more advantages than the l_p norm in sparse representation.Therefore,based on the previous work,an improved graph-based l_p norm semi-supervised dictionary learning is proposed by using the l_p norm.Experimental results on real data sets show that this algorithm is superior to other algorithms and norm semi-supervised dictionary learning algorithms based on graph structure.(3)In view of the non convex non smoothness of the model,we propose an effective algorithm based on the Block Coordinate Descent(BCD)method.