A Survey on Knowledge Graph-Based Recommender Systems

• Information overload motivates recommender systems, while collaborative filtering's data sparsity and cold-start limitations have encouraged hybrid approaches using interaction and content similarity.Systems also use side information such as item attributes, reviews, and social networks.Source: §1 · p. 1
• Knowledge graphs connect users and items through entity relations, which can improve recommendation precision and provide explanations by tracing relation sequences.For example, a movie may be recommended because it shares a genre with one the user previously watched.Source: §1 · p. 1
• The survey classifies knowledge graph-based recommenders as embedding-based, path-based, or unified methods, and reviews how they use knowledge graphs for explainable recommendation.It also compares similarities and differences among the three methods.Source: §1 · p. 2
• The survey covers knowledge graph recommendation for movies, books, news, products, points of interest, music, and social platforms, and collects the evaluated datasets.It categorizes recent works by application.Source: §1 · p. 2

2.1 Knowledge Graphs

• A knowledge graph represents entities and their relations as triples, forming a heterogeneous network whose linked attributes and relations provide strong representational capacity.Each fact takes the form (head entity, relation, tail entity).Source: §2.1 · p. 2
• Knowledge graphs are grouped into cross-domain examples such as Freebase, DBpedia, YAGO, and NELL, and domain-specific examples such as Bio2RDF.Freebase had more than 3 billion facts and almost 50 million entities by 2015, while CN-DBPedia had over 16 million entities and over 220 million relations.Source: §2.1 · p. 2

2.2 Recommender Systems

• Recommender systems learn user and item representations, score candidate items for a user, and recommend unobserved items by ranking their preference scores.The task is described across applications including movies, music, points of interest, news, and education.Source: §2.2 · pp. 2–3
• Collaborative filtering uses interaction patterns, content-based filtering uses item attributes and past interactions, and hybrid methods combine techniques to overcome individual limitations.CF can use explicit ratings or implicit clicks and views; content-based methods match candidate items with a user's profile.Source: §2.2 · p. 3
• KG-based recommenders incorporate knowledge graphs as side information into collaborative filtering to improve recommendation performance amid interaction sparsity and cold start.Without interaction records, user-user and item-item similarities are difficult to determine; adding user or item side information can improve CF recommendations.Source: §2.2 · p. 3

3 OVERVIEW

• A HIN has multiple entity or relation types, while a KG is a directed graph of subject–relation–object facts and can be viewed as a HIN.Each KG edge links a head entity to a tail entity through a relation.Source: §3 · p. 3
• A meta-path defines a relation sequence in a HIN, while a meta-graph combines meta-paths; KGE represents entities and relations as low-dimensional vectors [65].The embeddings preserve graph properties such as semantic meaning or high-order proximity.Source: §3 · pp. 3–4
• The binary user-feedback matrix records observed user–item interactions, which may be implicit rather than preference signals; explicit ratings can also be available.Implicit behaviors include clicking, watching, and browsing; movie ratings on a one-to-five scale are an example of explicit feedback.Source: §3 · p. 4

4 METHODS OF RECOMMENDER SYSTEMS WITH KNOWLEDGE GRAPHS

• The section groups collected methods by how they use KG information and organizes their publication, KG-use approach, and adopted techniques in Table 3.It also considers how KG-based reasoning can support explanations for recommendations [97].Source: §4 · pp. 4–5

4.1 Embedding-based Methods

• Embedding-based methods encode graph information to enrich item or user representations; experiments report that structural knowledge can boost recommendation performance.Methods may use item-only or user-item graphs, refine embeddings, or jointly train recommendation with graph-related tasks, motivated by shared features between item and entity embeddings.Source: §4.1 · pp. 5, 7, 8

4.2 Path-based Methods

• Path-based methods exploit graph connectivity and meta-path similarities to enhance recommendations through user and item similarity.Variants use user-specific meta-path weights, richer meta-graphs, and both favored and disliked past items; RuleRec learns weighted, explicit KG rules that make its recommendation process explainable.Source: §4.2 · pp. 8–9
• RKGE encodes user-item paths with a recurrent network, while KPRN encodes entity and relation sequences and pools path-level preference scores.RKGE averages path hidden states; KPRN uses weighted pooling.Source: §4.2 · p. 9
• Path-based methods make recommendations interpretable through predefined meta-paths, explicit rules, or automatically mined salient paths that explain user-item connections.RuleRec generates recommendation rules from an external KG; deep-learning approaches can mine salient paths automatically.Source: §4.2 · p. 10

4.3 Unified Methods

• Unified methods combine KG entity and relation representations with connectivity, refining user or item embeddings by propagating information across multi-hop neighbors.They commonly use GNN architectures, and the propagation process can be interpreted as discovering preference patterns in the KG.Source: §4.3 · pp. 10, 12
• RippleNet [14] propagates preferences outward from users’ historical interests, while AKUPM uses self-attention over click-history neighbors and RCoLM jointly trains KG completion and recommendation.RCoLM assumes items share the same latent representation in both modules, enabling mutual enhancement.Source: §4.3 · p. 10
• KGCN [89] aggregates multi-hop KG neighbors inward to item representations, while KGCN-LS adds label smoothness to propagate user-interaction labels.KGCN samples a fixed number of neighbors to form a receptive field, supporting efficient, scalable learning.Source: §4.3 · p. 11
• KGAT propagates embeddings outward in the user-item graph, while KNI [93], IntentGC, and AKGE [96] model cross-side neighbors, user-related behaviors, or pair-specific subgraphs.AKGE uses an attention-based GNN over a sampled subgraph to filter less-related entities and mine high-order user-item relations.Source: §4.3 · p. 11

4.4 Summary

• Embedding-based methods incorporate KG entity and relation embeddings but ignore informative connectivity patterns and often provide few explanations; path-based and unified methods can provide reasons.Path-based methods discover path-level similarity or connective patterns, and unified methods explain recommendation through embedding propagation.Source: §4.4 · p. 12

Movie

• Movie recommendation commonly uses MovieLens or DoubanMovie, with KGs connecting movies through attributes or incorporating users and rating relations.MovieLens-100K, MovieLens-1M, and MovieLens-20M are common benchmarks; item graphs support collaborative filtering, while user-item graphs can use ratings as relations,,,,,,,,,,, [93],,,,.Source: §5 · p. 12

Book and Music

• Book datasets include Book-Crossing, Amazon-Book, DoubanBook, DBbook2014, and IntentBooks; Last.FM and KKBox support music recommendation.Book KGs commonly map books to entities in external databases, whereas DoubanBook’s attributes support a user-item graph,,,,,,, [93],,, [87],,,. Music graphs may use Freebase or Satori, or be built from dataset information,,,,, [87], [96],,,.Source: §5 · p. 12

Dataset overview

• Table 4 organizes KG-based recommender datasets across seven application scenarios.The survey also compares how works construct a KG within each scenario.Source: §5 · p. 12, Table 4 · p. 13

News

• For news recommendation, Bing-News provides clicks and titles, and KGs connect title entities through neighboring entities in Satori.KGs are used to capture logical relations among news items in a time-sensitive, condensed-content scenario,,,.Source: §5 · p. 13

Social Platform

• Social-platform applications include recommending celebrities on Weibo, offline meetings on MeetUp, and conferences using DBLP data.For Weibo, the approach combines sentiment links in a user-item graph with celebrity information from Satori.Source: §5 · p. 13

Dynamic Recommendation

• Dynamic Recommendation: dynamic graph networks could capture rapidly changing interests, while integrating other side information into a KG may extend this approach.GNN- and GCN-based models can be time-consuming to train and thus model static preferences; Song et al. combine long- and short-term interests from friends with dynamic graph attention.Source: §6 · p. 13

Multi-task Learning

• Multi-task Learning: jointly training recommendation with KG completion, KGE, or item-relation regulation can address missing facts, while other KG tasks may transfer useful knowledge.Prior work reports benefits from joint KG completion and recommendation,, as well as recommendation with KGE or item-relation regulation; entity classification and resolution are proposed transfer sources.Source: §6 · pp. 13–14

Cross-Domain Recommendation

• Cross-Domain Recommendation: transfer from data-rich source domains can help target domains, and incorporating user and item side information into interaction graphs may improve recommendations.PPGN combines users and products from different domains in one graph and outperforms SOTA significantly, but its user-item graph includes only interaction relations; Zhang et al. proposed a matrix-based method and Zhao et al. introduced PPGN.Source: §6 · p. 14

Knowledge Enhanced Language Representation

• Knowledge Enhanced Language Representation: applying knowledge-enhanced text representation to news and other text-based recommendation may improve representation learning and recommendation accuracy.Chen et al. used KG knowledge for short-text classification, and Zhang et al. incorporated Wikidata in ERNIE for relation classification; DKN simply concatenates text and entity embeddings rather than fusing them.Source: §6 · p. 14

Knowledge Graph Embedding Method

• Knowledge Graph Embedding Method: compare translation-distance and semantic-matching KGE methods across data sources, recommendation scenarios, and model architectures to guide method selection.The survey reports no comprehensive guidance on which KGE method to use under different conditions.Source: §6 · p. 14

User Side Information

• User Side Information: incorporating user networks and demographic information into KG-based recommenders is a promising direction alongside the more common use of item side information.Fan et al. separately represented a user-user social network and user-item interaction graph with a GNN, outperforming traditional CF-based recommenders with user social information; [96] integrated user relations into the graph and showed this strategy's effectiveness.Source: §6 · p. 14

7 CONCLUSION

• The survey examines how KG-based recommender approaches use knowledge graphs as side information to improve recommendation accuracy and interpretability.It describes knowledge graphs as containing fruitful information that supports accurate and explainable recommendations.Source: §7 · p. 14
• It introduces datasets used in different recommendation scenarios.Source: §7 · p. 14
• It identifies future research directions intended to promote development in the field.Source: §7 · p. 14