Enhanced Graph Diffusion Learning with Transformable Patching via Curriculum Contrastive Learning for Session Recommendation

doi:10.3390/electronics14102089

Return

Enhanced Graph Diffusion Learning with Transformable Patching via Curriculum Contrastive Learning for Session Recommendation

2025-05-21

DOI：10.3390/electronics14102089

Original

Original request for help

Save

Abstract

En 中文

The fusion modeling of intra-session item information representation and inter-session item transition pattern for session recommendation has shown performance advantages. However, existing research still suffers from the following challenges: (1) the time-varying effects of complex relationships between item transitions within sessions need to be deeply explored; and (2) the lack of effective representation for inter-session item transition patterns. To address these challenges, we propose a new session recommendation, named EGDLTP-CCL. Specifically, we first design a patch-enhanced gated neural network representation of session item transition patterns, which accurately captures the time-dynamically varying impacts of the complex relationships within sessions of item transitions through a designed transformer patching strategy. Then, we develop an energy-constraint-based graph diffusion model to capture the inter-session item transition patterns, which mitigates the problem of poor simulation of real inter-session item transition patterns by the introduction of an energy-constraint strategy for the graph diffusion model. In addition, patch-enhanced gated neural networks and energy-constrained graph diffusion models are treated as two different views in the contrastive learning framework. By introducing a curriculum learning strategy that explores how to effectively select and train negative samples in a contrastive learning framework, thereby deeply improving performance in contrastive learning task. Finally, we combine and jointly train the recommendation task and the curriculum learning contrastive learning task for optimization based on a multi-task learning strategy to further improve the recommendation performance. Experiments on real-world datasets show that EGDLTP-CCL significantly outperforms state-of-the-art methods.

Keywords:

graph diffusion

transformer

contrastive learning

recommendation