[Paper Review] Application of Knowledge Graphs to Provide Side Information for Improved Recommendation Accuracy

Yuhao Mao, Serguei A. Mokhov+1 more|arXiv (Cornell University)|Jan 7, 2021

Recommender Systems and Techniques15 references4 citations

TL;DR

This paper proposes a generic software framework that integrates knowledge graphs as side information to enhance recommendation system accuracy. By encoding product and user metadata—such as actors, directors, and user demographics—into a knowledge graph and leveraging it during model training, the framework improves recommendation performance, with up to 1% higher AUC and accuracy compared to baseline models.

ABSTRACT

Personalized recommendations are popular in these days of Internet driven activities, specifically shopping. Recommendation methods can be grouped into three major categories, content based filtering, collaborative filtering and machine learning enhanced. Information about products and preferences of different users are primarily used to infer preferences for a specific user. Inadequate information can obviously cause these methods to fail or perform poorly. The more information we provide to these methods, the more likely it is that the methods perform better. Knowledge graphs represent the current trend in recording information in the form of relations between entities, and can provide additional (side) information about products and users. Such information can be used to improve nearest neighbour search, clustering users and products, or train the neural network, when one is used. In this work, we present a new generic recommendation systems framework, that integrates knowledge graphs into the recommendation pipeline. We describe its software design and implementation, and then show through experiments, how such a framework can be specialized for a domain, say movie recommendations, and the improvements in recommendation results possible due to side information obtained from knowledge graphs representation of such information. Our framework supports different knowledge graph representation formats, and facilitates format conversion, merging and information extraction needed for training recommendation methods.

Motivation & Objective

• Address the challenge of insufficient user and item information in traditional recommendation systems.
• Provide a unified software framework to integrate diverse side information from knowledge graphs into recommendation pipelines.
• Enable scalable, extensible, and interoperable use of structured knowledge (e.g., entities, relations) to improve model generalization and interpretability.
• Support multiple knowledge representation formats and machine learning backends for broad research and deployment compatibility.
• Demonstrate the framework’s effectiveness through empirical evaluation on the MovieLens-1M dataset using a knowledge-aware recommendation model (MKR)

Proposed method

• Represent side information (e.g., movie genres, actors, directors, user age, occupation) as triples in a knowledge graph using RDF, RDFS, OWL, or Neo4j formats.
• Implement a modular software framework in Python that supports knowledge graph ingestion, format conversion, merging, and information extraction for training.
• Integrate the knowledge graph with a deep learning recommendation model (MKR) by enriching item and user embeddings with graph-structured side information.
• Use a multi-task learning objective in the MKR model to jointly optimize for rating prediction and knowledge-aware contrastive learning.
• Support real-time inference by caching pre-trained graph structures and model weights, and measure latency using timestamped pipeline execution.
• Enable extensibility via a planned Java API wrapper and support for multiple machine learning frameworks (e.g., PyTorch, Scikit-learn)

Experimental results

Research questions

• RQ1 How does integrating structured side information from knowledge graphs affect the accuracy of recommendation models?
• RQ2 What types of side information (e.g., user demographics, movie posters, actors, directors) contribute most significantly to improved recommendation performance?
• RQ3 Can a generic, extensible software framework effectively support diverse knowledge graph formats and machine learning backends in recommendation systems?
• RQ4 How does the framework perform in terms of real-time inference latency compared to established baselines?
• RQ5 To what extent does sparse side information (e.g., movie posters) degrade model performance, and can it be improved via feature extraction?

Key findings

• The inclusion of both user and movie side information (e.g., age, occupation, actors, directors) increased test AUC to 0.9136, a 0.9% improvement over the baseline (0.9042).
• User-side information (age, job) contributed significantly to performance, with test AUC rising to 0.9091 when added to the baseline.
• Movie-side information (genres, actors, directors) alone improved test AUC to 0.9061, a 0.2% gain over the baseline.
• Poster-based side information had a negative impact due to data sparsity, reducing test AUC to 0.9029 and accuracy to 0.8214, suggesting limited utility without visual feature extraction.
• The framework achieved an average inference latency of 634.33ms on a laptop, well below the 2000ms real-time baseline, demonstrating practical deployability.
• The framework’s modular design supports multiple knowledge graph formats (RDF, OWL, Neo4j) and enables future integration of additional frameworks like PyTorch and Caffe.