Contrastive Pre-training for Sequential Recommendation
share
Sequential recommendation methods play a crucial role in modern recommender systems because of their ability to capture users' dynamic interests from their historical interactions. Despite their success, we argue that these approaches require huge amounts of parameters to learn a high-quality user representation model. However, they usually suffer from the data sparsity problem, which makes it difficult for them to collect sufficient supervised information to optimize the parameters. To tackle that, inspired by recent advances of pre-training techniques in the natural language processing area, we construct the training signal from unsupervised data and then pre-train the user representation model with this information. We propose a novel model called Contrastive Pre-training for Sequential Recommendation (CP4Rec), which utilizes the contrastive pre-training framework to extract meaningful user patterns and further encode the user representation effectively. In addition, we propose three data augmentation approaches to construct pre-training tasks and exploit the effects of the composition of different augmentations. Comprehensive experiments on four public datasets demonstrate that CP4Rec achieves state-of-the-art performance over existing baselines especially when limited training data is available.
READ FULL TEXT
