A newly published study from the Institute of Electrical and Electronics Engineers (IEEE) presents a novel approach to improving the accuracy and robustness of recommender systems, using a combination of knowledge graph embedding, two-dimensional convolutional models, and contrastive learning.
Addressing Cold-Start and Sparse Data Issues
The paper, titled “Research on the Application of Knowledge Graph-Driven Two-Dimensional Convolutional Embedding Methods in Recommender Systems”, was presented at the 2025 International Conference on Intelligent Systems and Computational Networks (ICISCN). The research offers an effective solution to long-standing issues in personalization technologies, including data sparsity and the cold-start problem.
- The study proposes a framework that extracts user preferences from interaction graphs and item features from knowledge graphs.
- A two-dimensional convolutional model, incorporating spatial and channel attention, is used to enhance feature extraction from the item side.
- Recommendation predictions are made through inner product operations between user and item vectors.
The system also introduces a contrastive learning strategy to further improve performance in real-world settings where data may be noisy or incomplete. By generating augmented subgraphs and comparing them to the original knowledge graph, the system learns hierarchical features that enhance both the accuracy and stability of recommendations.
“Real-world recommendation systems are often challenged by incomplete or noisy data,” said lead author Peng Dong. “Our goal was to build a model that not only improves accuracy, but also stays reliable under real-world constraints.”
Developing Two Models for Evaluation
The motivation behind the study led to the development of two models—KG-UIR and KGCRL—evaluated on multiple benchmarks. Results showed marked improvements in recommendation accuracy and a measurable reduction in the negative impact of cold-start and sparse-data scenarios.
| Model | Performance Metric | Benchmark |
|---|---|---|
| KG-UIR | Recommendation Accuracy | RM-100K |
| KGCRL | Recommendation Accuracy | RM-100K |
The research could be applied widely in industries where personalized user engagement is critical, including e-commerce, streaming media, digital advertising, and financial services. Systems built on this architecture could enable more relevant product suggestions, smarter content delivery, and more effective user targeting.
Author Background and Contact Information
Peng Dong, the paper’s lead author, holds a master’s degree in Business Intelligence and Analytics – Data Science Concentration from Stevens Institute of Technology. He has worked on large-scale recommendation and predictive modeling systems at companies such as Paramount and The Trade Desk. His broader research spans areas including interpretable AI, financial risk modeling, and the impact of corporate ownership structures on long-term financial sustainability.
For more information, please contact Peng Dong via email or visit his website at https://scholar.google.com/citations?hl=en&user=9lstaOsAAAAJ.
Conclusion
The proposed framework and contrastive learning strategy offer a promising solution to the challenges faced by recommender systems in handling incomplete or noisy data.
