CoFiGAN: Collaborative filtering by generative and discriminative training for one-class recommendation

摘要

In this paper, we study an important collaborative filtering problem with users’ one-class feedback such as purchases and likes that are pervasive in various recommendation scenarios. In particular, we make a significant extension of IRGAN by introducing rich interactions between a generator and a discriminator, and then design a novel collaborative filtering algorithm termed as CoFiGAN. In our CoFiGAN, the complementarity of the generative training and the discriminative training is exploited more completely, which enhances the accuracy of modeling users’ behaviors. Similar to other GAN-based algorithms, our CoFiGAN can also be interpreted as playing a minimax game, i.e., the generator generates samples close to the true ones aiming to confuse the discriminator, while the latter focuses on distinguishing between the true and generated samples. Different from others, the generator in our CoFiGAN generates items from a more direct and effective way under the guidance of the discriminator in order to accelerate convergence in adversarial training and increase the diversity of the generated samples to avoid mode collapse to some extent. Extensive empirical studies on four public and real-world datasets show that our CoFiGAN performs better than IRGAN and other very strong recommendation algorithms in terms of the commonly used ranking-oriented evaluation metrics.