Abstract
With the advancement of personalized education, group recommendation has become a core method for improving cooperative learning. However, existing methods often overlook the complex collaborative relationships between groups, members, and items. We propose Knowledge-aware Attentive Embeddings (KAEG), which utilizes a knowledge graph to capture structural information and interest similarities. By employing Simplified Graph Convolutional (SGC) networks and attention-embedded aggregators, KAEG models group preferences more effectively. A joint training strategy further enhances performance by co-optimizing user and group recommendation tasks. Experimental results on real-world datasets show that KAEG outperforms state-of-the-art baselines.
Introduction
In the context of Massive Online Open Courses (MOOCs), students frequently form temporary groups for cooperative learning. Recommending appropriate educational resources to these groups is challenging due to two primary factors: the difficulty of aggregating individual interests into a collective preference and the inherent data sparsity of occasional groups (the "cold-start" problem). Traditional recommendation systems often fail to capture the semantic "collaborative relationships"—the underlying links between teachers, concepts, and courses—that are vital for predicting group needs when historical data is scarce.
The KAEG Framework
The proposed KAEG model addresses these challenges by integrating Knowledge Graphs (KG) with advanced neural architectures. The framework consists of three main components: KG integration, Simplified Graph Convolutional (SGC) networks, and dual-channel attentive aggregators.
Knowledge Graph and SGC Networks
KAEG constructs a heterogeneous KG where entities such as courses, videos, and concepts are linked by various relations. This provides a semantic foundation for connecting disparate educational items. To propagate information through this graph, the model employs SGC networks. Unlike standard Graph Convolutional Networks, SGC removes non-linear activations between layers, which improves computational efficiency and prevents "over-smoothing" while capturing high-order similarities between users and groups.
Dual-Channel Attentive Aggregators
To model group preferences accurately, KAEG uses a dual-channel approach:
- Member Aggregation: An attention mechanism calculates the specific influence of each member on a candidate item, recognizing that expertise varies across subjects.
- Group Preference: A separate embedding captures the collective identity or general goal of the group. An adaptive neural network then merges these channels into a unified group representation.
Joint Training Strategy
The model is trained using a multi-task learning approach that simultaneously optimizes for individual user-item interactions and group-item interactions. This joint training allows the two tasks to reinforce each other, leading to more robust embeddings even for users with limited group history.
Experimental Evaluation
The researchers evaluated KAEG on two real-world datasets: MOOCCube, a large-scale MOOC repository, and Book-Crossing, used to test generalizability. Synthetic groups of 2-5 members were created to benchmark performance.
Key Results
KAEG consistently outperformed state-of-the-art baselines (e.g., NCF, AGREE, KGCN) across multiple metrics:
- Recommendation Accuracy: On the MOOC dataset, KAEG achieved a Hit Rate (HR@5) of 0.9224 and an NDCG@5 of 0.7567, significantly surpassing the AGREE model.
- Cold-Start Resilience: In scenarios where training data was limited to 20%, KAEG showed a much higher improvement ratio over baselines, demonstrating the effectiveness of the KG in compensating for missing interaction data.
- Ablation Studies: Removing either the KG component or the joint training strategy resulted in significant performance declines, validating the importance of each architectural choice.
Conclusion and Future Directions
KAEG represents a significant step forward in educational group recommendation by effectively leveraging side information through Knowledge Graphs. While the results are promising, future work should address the limitations of synthetic groups by incorporating real-world group dynamics. Additionally, transitioning from a static to a dynamic KG and incorporating multi-behavior data (e.g., distinguishing between resource consumption and task completion) could further enhance the model's utility in live educational settings.
Summary updated