💻 computer science

A Study on the Intelligent Recommendation of Interdisciplinary Student Innovation and Entrepreneurship Projects Integrating Knowledge Graphs and Graph Neural Networks

This paper proposes the KG-GNN-IPR model, which integrates knowledge graphs and graph neural networks to enhance the accuracy, robustness, and explainability of interdisciplinary student innovation and entrepreneurship project recommendations by effectively leveraging heterogeneous educational data and multi-hop skill complementarity.

Original authors: Yanhong Wu¹, Lina Guo¹, Qi Li¹, Chengyu Sun

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yanhong Wu¹, Lina Guo¹, Qi Li¹, Chengyu Sun

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern university, the most exciting work often happens where different fields collide. A student might need to combine computer coding with business strategy, or blend design thinking with agricultural science, to solve a real-world problem. These interdisciplinary projects are the heart of innovation and entrepreneurship education, yet finding the right team members for them is notoriously difficult. Traditionally, this matching process has relied on students browsing long lists of project titles or asking teachers for personal advice. While these methods work in small groups, they struggle when the number of students and projects grows, often overlooking how a student's specific course history, hidden skills, and past competitions connect to a project's complex needs. To solve this, researchers are turning to a digital approach that treats education not just as a list of grades, but as a vast, interconnected web of relationships.

A team of researchers at Shandong Huayu University of Technology has developed a new way to navigate this web. They built a system that acts like a detailed map of the entire educational landscape, connecting students to the courses they took, the skills they learned, the mentors who guide them, and the specific projects they might join. This map, known as a knowledge graph, allows the computer to see links that a human might miss, such as how a class on data analysis prepares a student for a project requiring machine learning, even if the student has never applied for a similar project before. By using a type of artificial intelligence designed to understand these complex connections, the researchers created a tool that can suggest the right projects to the right students with much greater accuracy than previous methods.

The researchers started by gathering real, anonymous records from their university's project management platform. They organized data on 2,000 students and 500 projects, linking them to 180 different courses, 120 specific skills, 90 mentors, and 12 major fields of study. They then constructed a massive network where every student, course, skill, and project was a point on the map, and the relationships between them were the lines connecting those points. For example, the system understood that a student who completed a Python course possessed a "programming" skill, and that a specific project required that exact skill. This structure allowed the computer to trace a path from a student's past learning directly to a project's future needs, even if they were separated by several steps in the network.

To test if this approach worked, the team compared their new system against several existing methods used in education and business. These older methods included simple popularity lists, systems that just matched students with similar past behaviors, and other advanced computer models that did not use this detailed map of relationships. The results showed that the new system, which combined the knowledge map with a smart learning algorithm, consistently outperformed all the others. When the researchers asked the system to recommend the top ten projects for a student, it was significantly more accurate at picking the ones the student would actually find relevant. The system was particularly good at handling difficult situations, such as when a student was new to the platform with no history, or when a project was brand new with no past applicants. In these cases, where other systems failed to make a connection, the new model could still find a match by looking at the student's courses and the project's required skills.

One of the most important findings was how well the system handled projects that crossed traditional boundaries. Interdisciplinary projects are often the hardest to fill because they require a mix of talents that do not usually appear together in a single student's record. The researchers found that their system was far better at identifying these cross-disciplinary matches than any of the older methods. It could see, for instance, that a marketing student with data analysis skills would be a perfect fit for a technology project, even though their major was different. The system did not just guess; it could show exactly why a match was made. It could trace a path showing that a student took a specific course, which taught a specific skill, which was then required by a specific project. This ability to explain its reasoning is crucial, as it allows teachers and students to trust the suggestions rather than viewing them as a mysterious black box.

The study also tested how robust the system was when the data was messy or incomplete, which is common in real-world university records. The researchers intentionally introduced errors, such as mislabeled skills or missing project requirements, to see if the system would break. Even with up to thirty percent of the data corrupted, the new system remained stable and accurate, while other methods struggled significantly. This resilience suggests that the network of relationships the system built was strong enough to compensate for missing or incorrect information. By looking at the broader picture of a student's education and a project's context, the system could find the right connections even when some pieces of the puzzle were slightly off.

Ultimately, this research suggests that the future of student project matching lies in understanding the deep, structural relationships within education, rather than just looking at surface-level similarities. The system does not replace the judgment of teachers or the choices of students; instead, it provides a powerful tool to highlight opportunities that might otherwise remain hidden. By making the invisible links between learning and application visible, the researchers have created a way to help students find the challenges that will best stretch their abilities and help them grow. The work demonstrates that when artificial intelligence is built on a clear map of how knowledge connects, it can offer guidance that is not only accurate but also understandable and deeply relevant to the human experience of learning.

Technical Summary: Intelligent Recommendation of Interdisciplinary Student Innovation and Entrepreneurship Projects

Problem Statement
University innovation and entrepreneurship education increasingly requires students to collaborate across disciplinary boundaries to solve open-ended practical problems. However, current project matching mechanisms on university platforms rely heavily on manual browsing, keyword screening, or teacher intuition. These traditional methods fail to capture the complex, multi-dimensional relationships between student attributes (course history, skills, interests), project requirements (technical routes, mentor expertise, disciplinary needs), and the structural nature of interdisciplinary teams. Furthermore, existing recommender systems often struggle with data sparsity, cold-start scenarios (new students or new projects), and the lack of explainability required for educational decision-making.

Methodology: The KG-GNN-IPR Model
The authors propose the KG-GNN-IPR model, which integrates Knowledge Graphs (KG) and Graph Neural Networks (GNN) to address these challenges. The methodology consists of four core components:

Heterogeneous Educational Knowledge Graph Construction:
The system constructs a graph $G=(V, E, R)$ containing eight entity types: students, courses, skills, majors, disciplines, mentors, projects, and competitions. Edges represent specific educational relationships, such as "completed course," "possesses skill," "project requires skill," and "mentor guides project." This structure allows the model to represent interdisciplinary links that are not captured by simple user-item interaction matrices.
Knowledge Graph Embedding:
To ensure semantic consistency, the model employs translational relation constraints (similar to TransE) where the sum of the head entity and relation embeddings approximates the tail entity embedding ( $\mathbf{e}_h + \mathbf{e}_r \approx \mathbf{e}_t$ ). This preserves the basic semantic structure of educational entities before graph propagation.
Relation-Attention Propagation (GNN):
The core of the model utilizes a Graph Attention Network (GAT) mechanism to aggregate information from neighbors. Unlike standard GCNs that average neighbor information, this module learns attention weights ( $\alpha_{vu}$ ) for different relation types. This allows the model to distinguish the importance of various paths (e.g., prioritizing "skill possession" over "major affiliation" depending on the context) and propagate signals across multiple hops to capture higher-order dependencies.
Matching and Ranking:
The final layer combines student and project representations using dot products, concatenation, and element-wise products to predict a matching score. The model is trained using Bayesian Personalized Ranking (BPR) loss, which optimizes the relative ordering of positive (participated) versus negative (non-participated) project pairs, rather than just predicting absolute scores.
Explainability Mechanism:
The system provides two levels of explanation:
- Path Interpretation: Tracing multi-hop paths (e.g., Student $\to$ Course $\to$ Skill $\to$ Project) to justify why a specific project is recommended.
- Attention Interpretation: Summarizing the contribution weights of different relationships to the final score.

Key Contributions

Domain-Specific Graph Construction: The study constructs a comprehensive heterogeneous graph specifically tailored to interdisciplinary innovation and entrepreneurship, explicitly modeling the interplay between skills, disciplines, mentors, and project stages.
Hybrid Architecture: It combines knowledge graph embedding for semantic stability with relation-aware GNN propagation for capturing complex, multi-hop structural patterns, addressing the limitations of both collaborative filtering and standard graph methods.
Robustness in Sparse and Cold-Start Scenarios: The model is designed to leverage side information (courses, skills, mentors) to make recommendations when historical interaction data is missing or sparse, a common issue in university project cycles.
Explainable AI for Education: The framework prioritizes interpretability, providing actionable paths and attention weights that allow educators and students to understand the rationale behind recommendations, aligning with human-centered educational data system design.

Experimental Results
The model was evaluated using de-identified data from Shandong Huayu University of Technology, comprising 2,000 students and 500 projects. It was compared against seven baselines: Popularity, UserCF, BPR-MF, GCN, GAT, LightGCN, and KGAT.

Overall Performance: KG-GNN-IPR outperformed all baselines across Precision@10, Recall@10, NDCG@10, and HR@10. Notably, it achieved an NDCG@10 of 0.444 and HR@10 of 0.756.
Cold Start & Sparsity: The model demonstrated superior resilience in cold-start scenarios (new students/projects) and under high data sparsity (10%–20% interaction data), where traditional matrix factorization and standard GNNs showed significant performance drops.
Noise Robustness: Under noise perturbations (up to 30% in skill labels and interaction edges), KG-GNN-IPR maintained higher performance than LightGCN, suggesting that the heterogeneous graph structure provides fault tolerance.
Interdisciplinary Performance: The model showed the smallest performance gap (-0.020) between single-discipline and interdisciplinary project recommendations, indicating its effectiveness in handling cross-domain skill matching.
Ablation Studies: Removing the knowledge graph, GNN propagation, or attention mechanism resulted in significant performance declines, confirming the necessity of each component.

Significance and Claims
The paper claims that the integration of knowledge graphs and GNNs offers a traceable semantic structure for educational entities, enabling the system to learn multi-hop links and patterns of skill complementarity that simple similarity matching cannot identify. The authors assert that this approach improves both the accuracy and explainability of recommendations, making it suitable for practical deployment in educational platforms.

The study emphasizes that the model serves as a decision support tool rather than a replacement for human judgment. By visualizing recommendation paths and highlighting ability gaps, the system aims to support teacher decision-making, student self-regulation, and the formation of effective interdisciplinary teams. The authors conclude that while the model shows promise, future work requires multi-university validation, dynamic graph updates to reflect student growth, and the integration of large language models for automated requirement extraction.

Technical Summary: Intelligent Recommendation of Interdisciplinary Student Innovation and Entrepreneurship Projects

More like this