Contrastive Generator Generative Adversarial Networks for Sequential Recommendation

Generally, sequential recommendation refers to the task of training a sequential generative model that can generate items of interest for users based on their historical interactions. However, these methods have some drawbacks as they require an effective generative model and training procedure to produce satisfactory results. These disadvantages include: 1) generative models lacking better generators for generative item sequences that user may be interested in, and 2) not distinguishing between true and fake sequences to capture users' preferences. In this paper, we propose Contrastive Generator Generative Adversarial Networks (CoGGAN) for sequential recommendation. We use a Residual AutoEncoder in Generator to generate item sequences, leveraging the powerful learning capacity of neural networks.Then Contrastive Learning is employed to obtain user preferences between the generative sequence and fake sequence. Additionally, we enhance the Discriminator by combining the Wasserstein loss with a ranking loss, creating a joint loss function.This combination better distinguishes between generated sequences and ground truth data, guiding the Generator towards producing item sequences that align more closely with user preferences. Extensive experiments demonstrate that CoGGAN significantly outperforms existing generative models across several publicly available datasets.