A Diffusion Data Enhancement Retentive Model for Sequential Recommendation
Tengqing Wu · 2024
When using deep learning techniques to process sequential recommendation content, neural network models are typically trained relying on a wealth of historical interaction feedback. However, this feedback data often contains natural or artificial noise, which can impact the extraction of interaction feature data by the model. Diffusion models, known for their powerful denoising capabilities, gain broad usage across image and natural language processing domains. Due to this powerful advantage, diffusion models theoretically have great potential for sequence-based recommendation tasks. In this article, we propose a diffusion data enhancement retentive model called DIDER for sequential recommendation. By applying a diffusion model during the denoising phase of neural networks, the model effectively enhances data features. Subsequently, a retentive model is introduced into sequential recommendation to leverage its powerful data processing capabilities for learning data features. Finally, the integration of acquired features is achieved through the application of a Multi-Layer Perception. To demonstrate the effectiveness of the proposed model, in-depth experiments involving two authentic datasets are carried out, revealing that the DIDER model excels in comparison to various baseline methods.