GAN-Based One-Class Classification for Personalized Image Retrieval

One-class classification for a personalized image retrieval system is one of most important research issues in machine learning. However, the conventional one-class classification techniques can have an overfitting problem. Thus, in this paper, we propose a novel one-class classification technique using the framework of generative adversarial nets (GAN) for image data. First, the support model and one-class model are trained with only positive-class data by a minimax game. At the end of this learning process, the one-class model learns the features of positive-class data very well while reducing generation error. One of our important findings is that the negative-class data generated by the support model help the one-class model conceptually and experimentally reduce the generative error. Using CIFAR-10, we show that our proposed technique outperforms the conventional technique by ∼10% in terms of F1 measure.