AceKG: Academic Knowledge Graph

• AceKG is a large-scale academic knowledge graph featuring 3.13B RDF triples covering papers, authors, fields, venues, and institutes.
• It employs a consistent ontology with over 100 relation types and unique URIs, ensuring precise entity disambiguation and comprehensive multi-relational coverage.
• AceKG serves as a benchmark environment for tasks such as link prediction, community detection, and scholar profiling using advanced embedding and inference methods.

AceKG is a large-scale academic knowledge graph (KG) comprising 3.13 billion RDF triples derived from a consistent ontology designed for high-fidelity academic data mining. Its schema represents essential entities and their relations—spanning papers, authors, fields of study, venues, and institutes—with fine-grained resolution achieved via unique URIs, addressing long-standing issues of name ambiguity and insufficient multi-relational coverage. AceKG serves as both a foundational academic resource and a rigorous benchmark environment for algorithmic evaluation in knowledge representation, link prediction, scholarly community detection, and author profiling (Wang et al., 2018).

1. Ontology: Entity Classes, Relations, and Hierarchies

AceKG instantiates a directed, labeled multi-relational graph $G = (E, R, T)$ with $E$ as entities, $R$ as relation types, and $T \subseteq E \times R \times E$ as factual triples. The five disjoint top-level entity classes are:

$$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$

Each entity $e \in E$ is assigned a unique URI for unambiguous identification, enabling, for example, distinct representations for similarly-named authors (such as ace:7E7A3A69 vs. ace:7E0D6766 for two different "Jiawei Han" entities).

The relation vocabulary $R$ is extensive ($\sim$100 types) and includes:

• $\mathsf{paper\_is\_written\_by}$: encodes authorship,
• $\mathsf{paper\_cites\_paper}$: denotes citation,
• $E$0,
• $E$1,
• $E$2: defines a DAG over fields of study,
• and further relations such as $E$3 and $E$4.

Hierarchical, compositional, and numeric data is attached via relations such as $E$5 (for taxonomy) and unary predicates (for literals including citation counts and publication dates).

2. Scale, Statistics, and Dataset Partitions

AceKG encodes $E$6 triples and $E$7 entities, with explicit class counts:

Entity Class	Count
Papers	$E$8
Authors	$E$9
Fields of Study	$R$0
Venues	$R$1 (journals) + $R$2 (conferences)
Institutes	$R$3

AceKG-derived benchmark datasets include:

• AK18K for link prediction ($R$4, $R$5; train/validation/test counts: $R$6130,265/7,429/7,336),
• Six field-specific heterogeneous collaboration graphs (e.g., Biology, Computer Science, Economics, Medicine, Physics, and a union of five fields),
• A "Google-Scholar-venue" graph of 600K papers, 635K authors, 151 venues, and 2.37M edges.

All primary data originates from the in-house Acemap repository, normalized and deduplicated; disambiguation is ensured via systematic URI assignment. Preprocessing also normalizes literals and dates to facilitate large-scale, machine-compliant operations.

3. Entity Alignment and Rule-based Inference

Entity alignment integrates AceKG with prominent external bibliographic databases (IEEE, ACM, DBLP). Mappings are computed by matching paper title and author lists, principally using cosine similarity of TF-IDF representations:

$R$7

Titles may also be matched using edit distance. At threshold $R$8, the mapping achieves coverage of 2.33M (IEEE), 1.91M (ACM), and 2.27M (DBLP) papers.

Rule-based inference employs Horn-style rules to further augment the graph. Representative rules include:

• $R$9,
• $T \subseteq E \times R \times E$0.

This suggests that AceKG provides both explicit and inferred relational knowledge, improving its utility for downstream analytics.

4. Benchmark Tasks: Link Prediction, Community Detection, and Scholar Profiling

AceKG supports systematic evaluation via canonical academic data mining tasks:

4.1 Link Prediction (Knowledge Embedding)

Given a triple set $T \subseteq E \times R \times E$1, entities $T \subseteq E \times R \times E$2 are embedded as $T \subseteq E \times R \times E$3 and relations $T \subseteq E \times R \times E$4 as $T \subseteq E \times R \times E$5, using scoring functions $T \subseteq E \times R \times E$6 to predict missing links. Evaluated methods and representative scoring functions include:

• TransE: $T \subseteq E \times R \times E$7,
• TransH: projection onto relation-specific hyperplanes,
• DistMult: $T \subseteq E \times R \times E$8,
• HolE: $T \subseteq E \times R \times E$9 (circular correlation),
• ComplEx: $$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$0 with complex-valued embeddings.

All models use embedding dimension $$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$1, margin $$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$2, learning rate $$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$3, and SGD with early stopping on validation MRR; loss $$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$4 is a margin ranking loss with one-entity corrupted negatives per triple.

4.2 Community Detection and Scholar Classification

On extracted collaboration graphs, network representation learning is performed via:

• DeepWalk: homogeneous random walks with Skip-Gram,
• LINE: first- and second-order proximity preservation,
• PTE: joint text-network embeddings,
• metapath2vec: heterogeneous walks along user-defined metapaths.

Resulting node embeddings feed into logistic regression (5-fold CV) for classification (metrics: Micro-F1, Macro-F1) and k-means clustering (NMI metric) for community detection.

5. Experimental Results and Analytical Insights

5.1 Link Prediction Performance on AK18K

Filtered MRR and Hits@$$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$5 (for $$\mathcal{C} = \{\, \mathsf{Paper},\; \mathsf{Author},\; \mathsf{Field},\; \mathsf{Venue},\; \mathsf{Institute}\, \}$$6) for several knowledge embedding models:

Model	MRR	Hits@1	Hits@10
TransE	0.719	62.7%	89.2%
TransH	0.701	61.0%	84.6%
DistMult	0.749	68.7%	86.1%
HolE	0.864	83.8%	88.2%
ComplEx	0.817	75.4%	89.0%

HolE and ComplEx exhibit superior capability for modeling antisymmetric relationships, which characterize all seven AK18K relation types. TransE outperforms TransH in Hits@10, plausibly due to a high proportion of many-to-many relations (94%) and a limited relation schema, where the benefit from hyperplane projection (TransH) is minimal. AK18K results fall between FB15K and WN18 benchmarks; this is attributed to moderate graph sparsity and the relatively simple relation schema.

5.2 Scholar Classification and Clustering

Micro-F1 and Macro-F1 scores (classification) and normalized mutual information (clustering):

Dataset	DeepWalk	LINE(1+2)	PTE	metapath2vec
biology	.792/.547	.722/.445	.759/.495	.828/.637
computer sci.	.545/.454	.633/.542	.574/.454	.678/.570
economics	.692/.277	.717/.385	.654/.276	.753/.485
medicine	.663/.496	.701/.577	.694/.555	.770/.659
physics	.774/.592	.779/.640	.723/.571	.794/.635
5-Fields union	.731/.589	.755/.655	.664/.528	.831/.682
Google-venues	.948/.942	.955/.949	.966/.961	.971/.968

Clustering (NMI):

Method	FOS	Google
DeepWalk	0.277	0.394
LINE(1+2)	0.305	0.459
PTE	0.153	0.602
metapath2vec	0.427	0.836

Metapath2vec, by exploiting heterogeneous context, consistently achieves superior classification and clustering performance. The Google-venue graph's higher data quality (top venues, reduced interdisciplinarity) yields increased performance across methods. A pronounced Micro-F1 to Macro-F1 drop indicates major label imbalance and cross-field scholarly activity.

6. Applications and Prospective Research Avenues

AceKG underpins multiple academic data mining tasks:

• Cooperation Prediction: Embeddings encode institutional, field, and citation proximity, supporting co-authorship link prediction.
• Author Disambiguation: Hetero-typed relational context enables distinction of same-name authors; iterative URI refinement is supported via embedding-space clustering.
• Rising-Star Detection: Temporal AceKG snapshots, combined with feature trajectories and embedding novelty, support early-career impact forecasting.

Future research directions include dynamic KG updates (streaming ingestion), fine-grained event extraction, graph neural networks over heterogeneous schemas, and cross-modal knowledge integration (e.g., PDF content, code repositories). The scale, ontological clarity, and connection to external bibliographic resources position AceKG as a foundational asset for scalable, multi-facet academic knowledge discovery and benchmarking (Wang et al., 2018).