Class-driven concept factorization for image representation

Recently, concept factorization (CF), which is a variant of nonnegative matrix factorization, has attracted great attentions in image representation. In CF, each concept is modeled as a nonnegative linear combination of the data points, and each data point as a linear combination of the concepts. CF has impressive performances in data representation. However, it is an unsupervised learning method without considering the label information of the data points. In this paper, we propose a novel semi-supervised CF method, called class-driven concept factorization (CDCF), which associates the class labels of data points with their representations by introducing a class-driven constraint. This constraint forces the representations of data points to be more similar within the same class while different between classes. Thus, the discriminative abilities of the representations are enhanced in the image representation. Experimental results on several databases have shown the effectiveness of our proposed method in terms of clustering accuracy and mutual information.