Construction and Application of an Automotive Knowledge Graph for Recommendation Systems

Background Knowledge graphs, first introduced by Google in 2012, have become a core AI technology applied in search, recommendation, advertising, risk control, and many other fields. The rapid development of AI has expanded their use to domains such as automotive.

Development Status Major internet companies (Facebook, Baidu, Alibaba, Tencent, Meituan) have launched their own knowledge graphs, demonstrating the technology’s maturity and business impact.

Goals and Benefits While most existing domain graphs focus on e‑commerce, medical, or finance, this work targets the automotive sector, defining entities like car series, models, dealers, manufacturers, and brands, and outlines a step‑by‑step construction methodology that supports cold‑start, recall, ranking, and display in recommendation pipelines.

Graph Construction Challenges The main difficulties include defining a schema without a unified standard, handling heterogeneous data types (structured, semi‑structured, unstructured), requiring deep domain expertise, and ensuring data quality through knowledge fusion and manual verification.

Architecture Design The system is divided into three layers: construction layer (schema definition, data modeling, knowledge fusion), storage layer (graph database and indexing), and service layer (intelligent inference and query services). The chosen storage backend is JanusGraph , which offers open‑source licensing, Hadoop integration, high concurrency, and native support for the TinkerPop/Gremlin graph model.

Construction Steps

Ontology design – defining concepts, relations, and attributes using tools such as Protégé .

Knowledge acquisition – extracting structured data via ETL, semi‑structured data via clustering (BIRCH) and TF‑IDF/BERT features, and unstructured data via triple extraction using pre‑training models (DocBert, ERNIE‑DOC) and sparse‑attention transformers.

Knowledge ingestion – representing facts as RDF triples (S, P, O) and storing them in JanusGraph, with composite and mixed indexes (e.g., mgmt.buildIndex('byNameAndAgeComposite',Vertex.class).addKey(name).addKey(age).buildCompositeIndex() ).

Service design – providing unified query APIs that hide Gremlin details. Example Gremlin queries:

g.V().has('price',gt(8)).has('price',lt(12)).order().by('sales',desc).valueMap().limit(1) – finds the top‑selling car in a price range.

g.V(xiaoming).repeat(out()).times(2).valueMap() – two‑hop neighborhood of a user.

g.V(xiaoming).repeat(out().simplePath()).until(or(has('car','name','kaluola'),has('car','name','xuanyi'))).path().by('name') – paths from a user to recommended articles.

Knowledge Graph in Recommendation

Cold‑Start – KG‑enhanced neural collaborative filtering (KGNCF‑RRN) and meta‑learning framework MetaKG leverage high‑order relations to alleviate data sparsity.

Explanation – path‑based explanations and the ECR multi‑task framework generate user‑friendly reasons for recommendations.

Ranking – KGAT uses graph neural networks with attention to capture high‑order item connections; RippleNet propagates user interests over the KG without handcrafted meta‑paths.

Conclusion The article details the end‑to‑end process of building an automotive knowledge graph, discusses technical challenges, presents concrete solutions (ontology design, data extraction, graph storage, and query services), and demonstrates its impact on recommendation scenarios such as cold‑start, explainability, and ranking.