Source-linked AI summary
AceKG: A Large-scale Knowledge Graph for Academic Data Mining
Ruijie Wang, Yuchen Yan, Jialu Wang, Yuting Jia, Ye Zhang, Weinan Zhang, Xinbing Wang
TL;DR
Existing academic knowledge graphs can lack multi-relational information and use ambiguous entity names. AceKG provides a large academic graph with a consistent ontology, database alignment, and rule-based inference. On AK18K, model performance varies by metric: TransE leads hit@10, while HolE and ComplEx perform best on other reported metrics.
Problem
Existing academic knowledge graphs may omit relations across entity types, focus on one field, or use entity names as identifiers despite synonymy and ambiguity.
Source: §1 · p. 1Method
AceKG builds a heterogeneous academic knowledge graph of 3.13 billion triples and enriches it through database alignment and rule-based inference.
Source: §1 · p. 1, §2.1 · p. 2, §2.3 · p. 2Results
On AK18K, TransE achieves 89.2% hit@10, while HolE and ComplEx achieve the strongest results on other reported metrics, including hit@1 and filtered MRR.
Source: §3.3 · p. 3Takeaways & Limitations
AceKG offers a large heterogeneous academic graph and benchmark setting for data-mining experiments and further research topics.
Source: §6 · p. 4
Abstract
from arXiv · showhide
Most existing knowledge graphs (KGs) in academic domains suffer from problems of insufficient multi-relational information, name ambiguity and improper data format for large-scale machine processing. In this paper, we present AceKG, a new large-scale KG in academic domain. AceKG not only provides clean academic information, but also offers a large-scale benchmark dataset for researchers to conduct challenging data mining projects including link prediction, community detection and scholar classification. Specifically, AceKG describes 3.13 billion triples of academic facts based on a consistent ontology, including necessary properties of papers, authors, fields of study, venues and institutes, as well as the relations among them. To enrich the proposed knowledge graph, we also perform entity alignment with existing databases and rule-based inference. Based on AceKG, we conduct experiments of three typical academic data mining tasks and evaluate several state-of- the-art knowledge embedding and network representation learning approaches on the benchmark datasets built from AceKG. Finally, we discuss several promising research directions that benefit from AceKG.
1 INTRODUCTION
Existing academic knowledge graphs lose cross-entity relations, limit cross-field study, and may identify entities by ambiguous names. AceKG addresses these issues with a large, heterogeneous academic graph organized as machine-readable triples.
1 INTRODUCTION
- Existing academic graphs often omit relations between different entity types, focus on a single field, or use ambiguous entity names as identifiers.These limitations constrain cross-field discovery and knowledge mining.Source: §1 · p. 1
- AceKG contains 3.13 billion academic-fact triples under a consistent ontology spanning papers, authors, fields of study, venues, institutes, and their relations.The paper also reports nearly 100G disk size.Source: §1 · p. 1
- AceKG is a heterogeneous academic network with multiple entity categories and relation types, supporting varied academic data-mining experiments.The paper presents this as an advantage over existing open academic KGs or datasets.Source: §1 · p. 1
- AceKG maps computer-science papers to ACM, IEEE, and DBLP data, helping researchers integrate information across databases.The introduction also describes entity mapping to computer-science databases.Source: §1 · p. 1
- AceKG represents information as structured triples, making it machine-readable and easy to process.Source: §1 · p. 1
2 THE KNOWLEDGE GRAPH
AceKG is a large academic knowledge graph whose ontology represents academic entities and their relations. The paper describes its entity identifiers, database alignment, and rule-based inference for adding relations.
2 THE KNOWLEDGE GRAPH
- AceKG is a large academic graph with 3.13 billion triples, covering almost the whole academic area and offering a heterogeneous academic network.The dataset is collected from Acemap.Source: §2 · p. 1
2.1 Ontology
- The ontology defines five academic entity classes—Papers, Authors, Fields of study, Venues, and Institutes—and represents facts as triples.Each entity receives a URI, allowing same-name scholars such as Jiawei Han to remain distinct.Source: §2.1 · pp. 1–2
2.2 Entity alignment
- AceKG aligns its computer-science papers with IEEE, ACM, and DBLP records, and the graph is periodically updated with the latest academic information.The paper reports mapping statistics in Table 2.Source: §2.2 · p. 2, Table 2 · p. 2
2.3 Inference
- Selected rule-based inference rules define new relations in AceKG, providing more comprehensive ground truth.Source: §2.3 · p. 2
3 KNOWLEDGE EMBEDDING
This section introduces AK18K, an AceKG-derived benchmark for link prediction, and evaluates five knowledge-embedding models. Results vary by metric and reflect the benchmark’s sparse, simple relation structure and large candidate sets.
3.1 Task Definition
- Knowledge embedding maps entities to k-dimensional vectors and scores triple plausibility; the section evaluates TransE, TransH, DistMult, ComplEx, and HolE on link prediction.The task predicts a missing entity given another entity and the relation in a latent triple.Source: §3.1 · p. 2
3.2 Experimental Setup
- AK18K is built from influential papers at 68 international venues, with author, field, and institute triples added before random train, validation, and test splits.The benchmark has only seven relation types and is sparser than FB15K but denser than WN18 by #Trip/#E.Source: §3.2 · p. 2
3.3 Evaluation Results
- 89.2% Hits@10 goes to TransE, while HolE and ComplEx lead on other metrics, including Hits@1 (83.8% and 75.4%) and filtered MRR (0.482 and 0.440).The authors relate HolE and ComplEx’s results to the antisymmetric relations in AK18K.Source: Table 4 · p. 3, §3.3 · p. 3
- On AK18K, results are lower than on WN18 but higher than on FB15K; the authors cite limited relation types and many possible objects per relation as constraints.For example, a paper-in-field relation may have thousands of possible objects per triple; results are compared with those reported in.Source: §3.3 · p. 3
4 NETWORK REPRESENTATION LEARNING
This section evaluates network representation learning on AceKG academic networks through scholar classification and clustering. metapath2vec leads the reported comparisons, while classification results also vary across field-of-study and Google-labeled datasets.
4.1 Task Definition
- The authors use DeepWalk, PTE, LINE, and metapath2vec to learn node representations, then evaluate scholar classification and clustering.The task maps each vertex in a network with topology and node attributes to a d-dimensional representation.Source: §4.1 · p. 3
4.2 Experimental Setup
- The evaluation networks comprise five field-of-study networks, their integrated graph, and a larger network based on Google Scholar venue categories.151 of 160 venues across eight categories were matched to AceKG; paper labels use field or venue category, and scholar labels use the majority of their publication labels, with ties broken randomly.Source: §4.2 · p. 3
4.3.1 Classification.
- The Google-labeled dataset differs in performance from the field-of-study datasets, which the authors attribute to different distributions of papers and scholars.The Google-labeled papers come from top venues, with few scholars active across categories; field-of-study datasets contain more cross-field papers and scholars.Source: §4.3.1 · p. 3
- Micro-F1 patterns are described as reflecting interdiscipline: Biology subfields are most independent, while CS subfields cross most; a Micro-F1-to-Macro-F1 decline, especially in Economy, indicates subfield imbalance.The authors interpret the highest and lowest Micro-F1 values as indicating the respective patterns.Source: §4.3.1 · p. 3
- metapath2vec performs significantly better than other methods on scholar classification, while DeepWalk and LINE achieve comparable performance.The authors suggest limited edge and node types may help methods for homogeneous networks learn comprehensive representations.Source: §4.3.1 · p. 3, Table 6 · p. 4
4.3.2 Clustering.
- metapath2vec outperforms the other models on clustering, evaluated with normalized mutual information (NMI).The experiments use k-means on the same node representations as classification and report averages over 10 runs.Source: Table 7 · p. 4, §4.3.2 · p. 4
5 FUTURE DIRECTIONS
AceKG is proposed as a resource for future academic-mining tasks, including cooperation prediction, author disambiguation, and finding rising stars.
Cooperation prediction
- Network representation learning results may support cooperation prediction by representing researcher features beyond obvious factors such as prior cooperators, neighborhoods, and citation relations.The passage notes that academic knowledge graphs may include features such as shared institutions or fields that existing factors can miss.Source: §5 · p. 4
Author disambiguation
- AceKG’s network structure and node attributes can enhance author disambiguation, with iterative disambiguation and graph-quality improvement proposed as a feedback loop.The passage also says author-disambiguation algorithms with good performance can be applied to AceKG.Source: §5 · p. 4
Finding rising star
- Rising-star prediction is framed as useful for universities and companies hiring young faculty or scientists, with prior approaches using publication growth, mentoring relations, and other factors.Source: §5 · p. 4
6 CONCLUSION
The conclusion presents AceKG as a large academic knowledge graph and benchmark resource. It reports three experimental evaluations, discusses potential research topics, and says the graph will be maintained and updated.
6 CONCLUSION
- AceKG contains 3.13 billion academic-fact triples under a consistent ontology, covering commonly used properties and relations for papers, authors, fields, venues, and institutes.Source: §6 · p. 4
- The paper compares several state-of-the-art approaches through three experimental evaluations on AceKG.Source: §6 · p. 4
- The authors propose additional research topics that could benefit from AceKG and say they will maintain and update its coverage.Source: §6 · p. 4