Abstract
Modern recommendation systems often struggle with data sparsity and the "cold start" problem. This paper presents a modular software framework designed to integrate Knowledge Graphs (KGs) into the recommendation pipeline. By leveraging side information—such as item metadata and user demographics—the framework enhances prediction accuracy. We detail an end-to-end architecture comprising data extraction, storage management, visualization, and a deep learning-based recommendation engine. Experimental results on the MovieLens 1M and Book-Crossing datasets demonstrate that incorporating multi-faceted side information significantly improves Area Under the Curve (AUC) and overall prediction performance.
Introduction
Recommendation systems are foundational to the digital economy, powering platforms like Netflix, Amazon, and YouTube. However, traditional collaborative filtering methods frequently suffer from inadequate information, particularly when user-item interaction data is sparse. This limitation is most acute during the "cold start" phase, where new users or items lack sufficient history for accurate modeling.
While "side information" (contextual data about users and items) offers a potential solution, its integration remains technically challenging. Many existing tools are either domain-specific or lack a comprehensive end-to-end pipeline for data processing. As shown in the comparison of existing recommendation frameworks, there is a clear need for a more flexible, format-agnostic architecture.
System Architecture
The proposed framework is built upon a software engineering pattern of "Frozen Spots" (reusable components) and "Hot Spots" (domain-specific specializations). This modularity allows researchers to adapt the pipeline to various domains with minimal friction. The architecture is divided into four primary functional blocks:
- InfoExtractor: A web-crawling module that utilizes predefined extraction rules to harvest data from external sources like IMDB or Wikipedia.
- StorageManager: A backend capable of handling diverse graph formats, including Resource Description Framework (RDF) and graph databases such as Neo4j.
- Knowledge Graph Viewer: A visualization component using libraries like
networkxandOpenGLto allow researchers to inspect entity relationships. - Recommendation System (RS) Module: The core engine, which utilizes a modified Multi-task Feature Learning for KG enhanced Recommendation (MKR) model implemented in TensorFlow and Keras.
Methodology and Datasets
The RS module treats recommendation and KG feature learning as dual tasks. By sharing parameters between these tasks, the system learns the underlying logic of item relationships—for instance, recognizing that two movies are linked by a shared director or genre.
To evaluate the framework, we utilized two benchmark datasets:
- MovieLens 1M: This dataset was augmented with side information (directors, writers, actors, and posters) crawled via the InfoExtractor from IMDB.
- Book-Crossing: A dataset containing book ratings and user demographics (location, age), used to test the framework's cross-domain versatility.
Key Results and Insights
Our experiments confirmed that the inclusion of side information leads to measurable gains in recommendation quality.
Accuracy Improvements
By integrating both item-side (movie) and user-side (demographics) information, the baseline model's AUC increased from 0.9042 to 0.9136. This represents a significant improvement in the system's ability to correctly rank user preferences. The inclusion of user context (age, occupation, gender) into the deep learning embeddings was particularly effective.
The Poster Paradox
One unexpected finding involved the use of visual data. When movie posters were added as side information, the AUC actually decreased to 0.9029. This "Poster Paradox" is attributed to data sparsity; unlike a director who might link dozens of films, a poster typically only connects to a single movie. Without dense relational links, the KG cannot effectively propagate features.
Operational Efficiency
Despite the computational overhead of processing complex KGs, the framework maintained high performance. The system demonstrated a response time of approximately 634ms, well within the 2000ms threshold required for real-time industry applications.
Conclusion and Future Work
This framework provides a robust foundation for KG-enhanced recommendations. Future iterations will focus on expanding the ecosystem, including the development of a Java API wrapper for production environments and support for additional machine learning backends like PyTorch and Scikit-learn. Furthermore, we intend to apply computer vision techniques to extract features from posters—such as color palettes or recurring actors—to transform isolated images into useful, interconnected nodes within the Knowledge Graph.
Summary updated