Source-linked AI summary
A Survey on Knowledge Graph-Based Recommender Systems
Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu Zhu, Xing Xie, Hui Xiong, Qing He
TL;DR
Recommender systems face information overload, data sparsity, and cold-start challenges. This survey reviews how knowledge graphs serve as side information for accurate and explainable recommendation, covering methods, datasets, applications, and future directions.
Problem
Recommender systems still face data sparsity and cold-start problems while modeling preferences amid information overload.
Method
The survey categorizes KG-based recommender algorithms as embedding-based, path-based, or unified, and reviews their use of KGs, datasets, and applications.
Results
The survey finds that KG-based systems support more accurate and explainable recommendation by leveraging structured relations and interpretable paths.
Takeaways & Limitations
Knowledge graphs are presented as promising side information for accurate and explainable recommendation across multiple application scenarios.
Takeaways & Limitations
The survey identifies time-consuming training for GNN- or GCN-based systems and missing KG facts as future research challenges.
Abstract
from arXiv · showhide
To solve the information explosion problem and enhance user experience in various online applications, recommender systems have been developed to model users preferences. Although numerous efforts have been made toward more personalized recommendations, recommender systems still suffer from several challenges, such as data sparsity and cold start. In recent years, generating recommendations with the knowledge graph as side information has attracted considerable interest. Such an approach can not only alleviate the abovementioned issues for a more accurate recommendation, but also provide explanations for recommended items. In this paper, we conduct a systematical survey of knowledge graph-based recommender systems. We collect recently published papers in this field and summarize them from two perspectives. On the one hand, we investigate the proposed algorithms by focusing on how the papers utilize the knowledge graph for accurate and explainable recommendation. On the other hand, we introduce datasets used in these works. Finally, we propose several potential research directions in this field.
1 INTRODUCTION
Recommender systems address information overload but face data sparsity and cold-start challenges. This survey examines knowledge graphs as side information for more accurate and explainable recommendation.
- Information overload makes it difficult for users to identify interesting choices across applications such as music, movies, and online shopping.
- Collaborative filtering is widely used to model preferences from interaction data but suffers from data sparsity and cold-start problems.
- Knowledge graphs connect users, items, and attributes through heterogeneous entities and relations, helping capture latent user-item relations.
- KG-based recommendation can improve precision and provide explanations by tracing relation sequences between users and recommended items.
- The survey categorizes KG-based recommender systems into embedding-based, path-based, and unified methods, and reviews their applications and datasets.
2 RELATED WORK
This section introduces knowledge graphs and recommender systems, then frames KG-based recommendation as collaborative filtering augmented with structured side information.
- 2.1 Knowledge Graphs: Knowledge graphs represent large-scale, multi-domain information as entities connected by relational links, enabling attributes and higher-level relations to be discovered.
- 2.1 Knowledge Graphs: The survey distinguishes cross-domain knowledge graphs, including Freebase, DBpedia, YAGO, and NELL, from domain-specific graphs such as Bio2RDF.
- 2.2 Recommender Systems: Recommender systems learn user and item representations, score candidate items, and recommend them by sorting preference scores.
- 2.2 Recommender Systems: Collaborative filtering uses explicit or implicit user-item interactions to infer preferences from similar users or items.
- 2.2 Recommender Systems: Sparse interaction data makes similarity estimation difficult, while cold start prevents reliable similarity determination for new users or items without interaction records.
- 2.2 Recommender Systems: KG-based recommender systems combine collaborative filtering with knowledge graphs as side information for more accurate recommendation.
3 OVERVIEW
The survey defines the graph structures, interactions, neighborhoods, and relevance concepts used to describe KG-based recommendation methods.
- Foundations: A heterogeneous information network is a directed graph with entity-type and link-type mappings, containing multiple entity or relation types.
- Foundations: A knowledge graph is a directed graph whose edges are subject-property-object triples connecting head entities to tail entities through relations.
- Foundations: A meta-path defines a relation sequence between entity types and extracts connectivity features from a heterogeneous information network.
- Interactions: An observed user-item interaction is represented by Rij = 1, while unobserved interaction is represented by 0.
- Interactions: Implicit feedback includes behaviors such as clicking, watching, and browsing, but does not necessarily indicate user preference.
- Graph Concepts: An H-hop neighbor is connected to a node through a multi-hop relation path, with the zero-hop neighborhood containing the node itself.
- Graph Concepts: A user’s k-hop relevant entities are defined from the interaction matrix and knowledge graph, while user and entity ripple sets organize related knowledge triples across hops.
4 METHODS OF RECOMMENDER SYSTEMS WITH KNOWLEDGE GRAPHS
The survey organizes KG-based recommender systems into embedding-based, path-based, and unified methods, examining how they support accurate and explainable recommendation. Embedding methods enrich representations with KG information, while path-based methods exploit graph connectivity and make recommendation reasoning more interpretable.
- KG-based recommender systems are grouped into embedding-based, path-based, and unified methods according to how they use KG information.The survey compares these categories and summarizes their publication information, KG utilization strategies, and adopted techniques.
- Embedding-based Methods: Embedding-based methods encode KG information into low-rank entity embeddings and use these representations to enrich user or item representations.Knowledge graph embedding algorithms include translation-distance models such as TransE and semantic-matching models such as DistMult.
- Embedding-based Methods: These methods may combine structural, textual, and visual item knowledge, then calculate preference scores by comparing learned user and item representations.CKE, for example, combines structural knowledge encoded with TransR with textual and visual information before ranking items by preference score.
- Embedding-based Methods: 54: incorporating structural knowledge can boost the performance of recommendation.
- Embedding-based Methods: Multi-task embedding methods jointly train recommendation and KG-related objectives because item embeddings share features with associated KG entity embeddings.The combined objective balances recommendation and KG losses with a hyperparameter λ.
- Path-based Methods: Path-based methods exploit connectivity patterns or semantic paths between users and items, using predefined or automatically mined paths to enhance recommendation.RuleRec learns explicit rules and weights from an external KG, while RKGE encodes user-to-item paths with a recurrent network.
- Path-based Methods: Path-based methods provide interpretability because recommendation reasons can be traced to explicit rules, meta-paths, or salient paths in the graph.Recent approaches automatically mine salient paths for user-item pairs rather than relying solely on manually defined meta-paths.
5 DATASETS OF RECOMMENDER SYSTEMS WITH KNOWLEDGE GRAPH
The survey organizes KG-based recommender-system datasets across application scenarios and shows how knowledge graphs are constructed for different recommendation tasks. This overview is intended to help researchers select suitable datasets for evaluation.
- Dataset overview: The dataset overview explains how knowledge graphs are constructed for recommendation tasks and helps researchers find suitable evaluation datasets.This is presented as a two-fold contribution of the dataset section.
- Dataset overview: Datasets are grouped into seven application scenarios, with Table 4 collecting the datasets and corresponding papers.The section compares scenarios and explains how each dataset supports KG construction for recommendation.
- Movie: Movie recommendation commonly uses MovieLens and DoubanMovie datasets, which contain movie, user, social, and attribute information.MovieLens provides stable benchmark datasets, while DoubanMovie includes social relations and user and movie attributes.
- Application scenarios: The survey uses datasets spanning different application settings, including movie, location-based, news, meeting, and academic recommendation.The cited dataset discussion illustrates that KG side information is incorporated across varied recommendation scenarios.
- Movie: Movie-centric item graphs connect movies through attributes such as genres, countries, actors, and directors.These graphs are constructed from external sources including Satori, DBpedia, Freebase, CN-DBPedia, and IMDB.
- Location-based recommendation: In location-based recommendation, datasets such as Yelp Challenge support user-item graphs built from check-ins, reviews, and attributes.The section also mentions item-graph construction and datasets for trip and restaurant recommendation.
6 FUTURE DIRECTIONS
The survey identifies future directions for KG-based recommendation, including dynamic preference modeling, knowledge transfer, cross-domain recommendation, and richer user-side information. These directions address changing interests, incomplete knowledge, domain imbalance, and underused user information.
- Dynamic Recommendation: Dynamic graph networks are proposed to model rapidly changing user interests that static preference models may miss.The discussion highlights online shopping, news, Twitter, and forums as settings where interests can change quickly.
- Knowledge transfer and completion: Jointly training knowledge-graph completion and recommendation modules is reported as effective when missing relations or entities impair preference modeling.The survey also points to multi-task learning with KGE and item-relation regulation tasks.
- Cross-Domain Recommendation: Cross-domain recommendation can transfer interaction data from richer source domains to sparser target domains.The survey motivates this direction by noting unequal interaction density across domains such as Amazon books and other categories.
- User Side Information: User networks and demographic information are identified as underused side information that can be integrated into KG-based recommender systems.The survey cites separate representations of user-user social and user-item interaction graphs, as well as integrated user relations.
7 CONCLUSION
The survey reviews how knowledge graphs are used as side information for accurate and interpretable recommendation and introduces datasets across application scenarios. It also identifies future research directions and concludes that KG-based recommender systems are promising.
- Survey scope: The survey summarizes recent KG-based recommender-system research and explains how different approaches use KGs as side information.Its focus is the relationship between KG use, recommendation results, and interpretability.
- Contributions: It introduces datasets used in different recommendation scenarios and identifies future research directions in the field.These components are presented as part of the survey’s overall contribution.
- Conclusion: The survey concludes that KG-based recommender systems are promising for accurate and explainable recommendation because KGs contain fruitful information.This conclusion stays within the survey’s stated scope.