Augmenting Short Sequence for Robust Session-based Recommendation

In recent years, recommendation systems leveraging deep neural networks have gained prominence as a means to address the information overload challenge exacerbated by the rapid advancement of the Internet. Predominantly, deep recommendation models focus on modeling user attribute features by utilizing the sequence of previously interacted items to predict the next item a user is likely to engage with. However, prior to user registration or login on an application, these recommendation models can only access a limited scope of user interactions, typically restricted to a short sequence of recent sessions. This brevity inherently constrains the accuracy of the recommendation model. To mitigate this limitation, we draw inspiration from the concept of sequence augmentation in contrastive learning, which employs short sequences to derive additional supervised signals based on association rules. Furthermore, we incorporate the notion of non-negative transfer learning to enhance the extraction of supervised signals that better align with the underlying intent of user sessions. Comprehensive experiments conducted on three real-world datasets substantiate the efficacy of our proposed model, demonstrating its superiority over traditional RNN, CNN, GNN, and various Transformer-based approaches.