The field of recommender systems and graph learning is rapidly evolving, with a focus on improving the accuracy and diversity of recommendations. Recent developments have seen the integration of graph neural networks and multi-objective retrieval frameworks to balance semantic relevance and user engagement. Item-centric exploration and radial neighborhood smoothing are also being explored to address the cold start problem and improve recommendation efficiency. Furthermore, non-parametric graph convolution and self-loop iterative fusion of heterogeneous auxiliary information are being used to enhance ranking and recommendation performance. The use of contrastive cascade graph learning and test-time fast adaptation is also being investigated for classifying real and synthetic information diffusion patterns and predicting multi-scale information diffusion. Additionally, researchers are working on unifying self-supervised and supervised learning for graph recommendation and developing frameworks for fairness and diversity in recommender systems. Noteworthy papers include:
- Balancing Semantic Relevance and Engagement in Related Video Recommendations, which introduces a novel multi-objective retrieval framework to enhance standard two-tower models.
- Item-centric Exploration for Cold Start Problem, which introduces the concept of item-centric recommendations to identify the optimal users for new items.
- SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation, which combines supervised and self-supervised learning into a cohesive supervised contrastive learning loss.