Introduction
In the current era of "information overflow," recommender systems (RS) have become indispensable tools for navigating massive datasets. However, traditional Collaborative Filtering (CF) methods often encounter accuracy bottlenecks. These systems frequently struggle to capture the nuanced, multi-layered relationships between users and items when relying on individual classifications. Furthermore, analyzing user behavior at scale using raw graphs is computationally cumbersome. This paper proposes a hybrid framework that integrates ensemble learning with graph embeddings to enhance recommendation accuracy and efficiency.
Methodology
The authors introduce a two-step framework designed to group similar users and model their behavior effectively.
1. Ensemble Learning and Fuzzy Classification
Rather than analyzing users in isolation, the system employs ensemble learning to categorize users into neighborhoods. The model defines four distinct types of similarity:
- SF: Similarity in film selection (ratings).
- SFS: Similarity in film subjects (genres).
- SMPF: Similarity in selecting popular films.
- SMPSF: Similarity in selecting films with popular subjects.
These metrics are processed through fuzzy rules and decision trees to identify the "top three" classes of users most similar to a target user based on both demographic profile (age, gender) and behavioral patterns.
2. Graph Embedding with node2vec
The behavior of these identified user groups is modeled as a heterogeneous knowledge graph. To ensure the graph is searchable and computationally efficient, the authors utilize node2vec. This technique transforms graph nodes into low-dimensional vectors (embeddings) while preserving their structural neighborhood relationships, allowing the system to discover subtle, multi-hop relationships invisible to standard matrix-based approaches.
3. User Triple Vector (UTV)
The final recommendations are generated via a "User Triple Vector." This component calculates the cosine similarity between the target user’s embedding and candidate movie embeddings to produce a ranked "top-k" list of recommendations.
Experimental Setup
The study utilizes two industry-standard benchmarks to evaluate performance: MovieLens 1M and MovieLens 10M. To simulate real-world conditions where user-item interaction data is often incomplete, the authors intentionally introduced sparsity into the datasets during testing.
Key Results and Discussion
The proposed UTV model consistently outperformed standard Collaborative Filtering (CF), Content-Based Filtering (CBF), and modern graph-based models such as RippleNet across several metrics:
- Root Mean Square Error (RMSE): On the MovieLens 1M dataset, UTV achieved an average RMSE of 1.676, significantly lower than RippleNet (2.108) and Genetic Algorithm-based models (2.475).
- Accuracy (ACC): On the 10M dataset, UTV reached an average accuracy of 0.3615, surpassing RippleNet’s 0.3078.
- Area Under Curve (AUC): In top-N recommendation tasks, UTV demonstrated superior performance, with AUC scores reaching up to 0.8355 on the 10M dataset as the user base scaled.
These results highlight the power of group-based classification. By aggregating users into fuzzy-defined neighborhoods first, the system filters out the noise that typically plagues standard CF methods. Additionally, the use of fuzzy logic better accommodates the ambiguity of human taste, such as the subjective difference between a 3-star and 4-star rating.
Conclusion and Future Work
This research demonstrates that combining ensemble learning with structural graph learning significantly improves recommendation quality. While the UTV model shows high efficiency, future research could address the "cold-start" problem for users with missing metadata and explore the computational costs of real-time model retraining as new data enters live systems. The authors suggest this methodology is highly adaptable for other social networks and IoT-based recommendation environments.
Summary updated