Knowledge Graph Embeddings

• Knowledge Graph Embeddings are machine learning methods that encode entities and relations as vectors or structured objects, enabling efficient inference and link prediction.
• They encompass diverse paradigms such as translational, geometric, and analytical models, employing various scoring functions like TransE and ComplEx to capture relational properties.
• Recent innovations integrate multi-modal data and continual learning approaches, enhancing robustness, interpretability, and scalability in practical applications.

Knowledge Graph Embeddings (KGE) are a family of machine learning methods that represent entities and relations in knowledge graphs as vectors or structured objects in low-dimensional spaces. The principal aim of these methods is to inject the symbolic, discrete relational knowledge encoded by triples $(h, r, t)$ into continuous representation spaces, thereby enabling efficient and scalable inference for tasks such as link prediction, entity classification, and multi-hop logical reasoning (Niu, 2024, Ge et al., 2023, Bianchi et al., 2020, Cao et al., 2022, Peng et al., 2021).

1. Foundations: Definitions, Objectives, and Evaluation

A knowledge graph (KG) is formally specified as $$\mathcal{G} = (\mathcal{E}, \mathcal{R}, \mathcal{T})$$, with entities $\mathcal{E}$, relations $\mathcal{R}$, and observed triples (facts) $$\mathcal{T} \subseteq \mathcal{E} \times \mathcal{R} \times \mathcal{E}$$. The KGE paradigm learns a mapping:

• entities $e \mapsto \mathbf{e} \in \mathbb{F}^d$
• relations $r \mapsto \mathbf{r} \in \mathbb{F}^d$ or higher-order tensors

where $\mathbb{F}$ is typically $\mathbb{R}$, $\mathbb{C}$, or a geometric algebra, and $d \ll |\mathcal{E}|$. Triple plausibility is scored via a parametric function $f(h, r, t)$, generally chosen so that true triples are scored higher (or closer in distance) than corrupted triples (Peng et al., 2021, Ge et al., 2023, Niu, 2024).

Evaluation metrics:

• Mean Reciprocal Rank (MRR) and Hits@K: For link prediction, these metrics are computed by ranking candidate entities for given $(h, r, ?)$ or $(?, r, t)$ queries, both in "raw" and "filtered" settings (Bianchi et al., 2020, Niu, 2024).
• Ablations and clustering/semantic criteria: For evaluation of interpretability and downstream utility (Hubert et al., 2023, Peng et al., 2021).

2. Core Model Families: Algebraic, Geometric, and Analytical Paradigms

The diversity of KGE models can be understood through the lens of mathematical representation spaces (Cao et al., 2022). The main taxonomies are:

A. Algebraic Models (Vector, Group, Ring Structures):

• Translational methods: Entities and relations as points in $\mathbb{R}^d$; relations modeled as translations ($h + r \approx t$). TransE is canonical (Ge et al., 2023, Bianchi et al., 2020).
• Group- and ring-based models: Extensions such as TorusE (entities as points on a torus, supporting translation modulo periodicity (Cao et al., 2022)) and MöbiusE use more exotic group operations.

B. Geometric Models (Curvature and Metric Space):

• Euclidean models: Classic translational/rotation approaches (TransE, RotatE) (Bianchi et al., 2020, Xiao et al., 2024).
• Hyperbolic models: MuRP, ATTH, and others utilize hyperbolic space, with Möbius addition and distance, directly modeling exponentially expanding hierarchies (Niu, 2024, Cao et al., 2022).
• Spherical models: Embeddings on spheres (SEA, HyperspherE) are suited for data with cyclic or symmetric relational structure.

C. Analytical Models (Probability and Manifold):

• Distributional approaches: KG2E embeds entities/relations as full distributions (e.g., multivariate Gaussians), scoring via KL-divergence (Cao et al., 2022, Xiao et al., 2024).
• Normalizing flow models: Recent architectures (NFE) represent entities/relations as invertible flows on random variables, unifying and extending point/vector methods (Xiao et al., 2024).

3. Canonical KGE Models and Scoring Functions

Major architectures and their scoring mechanisms include:

Model	Representation space	Key scoring function (for triple (h,r,t))
TransE	$\mathbb{R}^d$	$- \|\mathbf{h} + \mathbf{r} - \mathbf{t}\|_p$
DistMult	$\mathbb{R}^d$	$\langle \mathbf{h}, \mathbf{r}, \mathbf{t} \rangle$
ComplEx	$\mathbb{C}^d$	$$\Re\langle \mathbf{h}, \mathbf{r}, \overline{\mathbf{t}} \rangle$$
RESCAL	$\mathbb{R}^d$	$\mathbf{h}^\top W_r \mathbf{t}$
RotatE	$\mathbb{C}^d$	$- \|\mathbf{h} \circ \mathbf{r} - \mathbf{t}\|_2$
TuckER	Tensor core	$$\mathcal{W} \times_1 \mathbf{h} \times_2 \mathbf{r} \times_3 \mathbf{t}$$
GeomE	Clifford/GA	$\langle E_h R_r \widetilde{E_t} \rangle_0$

These models exhibit differing capabilities in capturing key relation properties: symmetry, antisymmetry, inversion, composition, mapping cardinalities, and hierarchical or cyclic patterns (Ge et al., 2023, Niu, 2024, Xu et al., 2022).

Advanced combinations (e.g., CompoundE, CompoundE3D) parameterize relations as sequences or compositions of geometric transformations (translation, rotation, scaling, reflection, shear) in higher-dimensional (including affine and 3D) spaces (Ge et al., 2023).

4. Recent Innovations: Continual, Multimodal, and Interpretable KGE

Continual and Incremental Learning:

Modern KGs evolve, necessitating dynamic adaptation of embeddings. Schema-informed initialization, leveraging ontological class centroids and variance, accelerates incremental updates, improves link prediction, and mitigates catastrophic forgetting across multiple backbone models (TransE family, RotatE, DistMult, HolE, etc.) (Pons et al., 14 Nov 2025).

Multimodal KGEs:

VL-KGE integrates pretrained vision-LLMs (CLIP, BLIP) and structural KGE backbones by fusing text, image, and learned entity representations. Fusion operators (average, weighted, concatenation-projection) enable robust inductive inference on entities with missing modalities and significantly outperform unimodal and prior multimodal methods in both symmetric and asymmetric MKGs (Efthymiou et al., 2 Mar 2026).

Interpretability and Semantic Grounding:

Interpretable KGE is pursued via:

• Disentanglement regularization (component-wise semantic isolation) (Kou et al., 2020).
• Schema/ontology pretraining (MASCHInE), where pretraining on "protograph" (class-level) structure produces embeddings better respecting domain/range constraints and yielding improved clustering and semantic validity in link prediction (Hubert et al., 2023).
• Normalizing flow-based KGE (NFE) unifies and extends prior models by casting embedding transformations as elements of the permutation group acting on random-variable spaces (Xiao et al., 2024).

5. Optimization, Training Efficiency, and Robustness

Standard optimization involves margin-based ranking losses or logistic objectives with negative sampling (Bianchi et al., 2020, Niu, 2024). Recent works emphasize computational and environmental efficiency:

• ProcrustEs segments entity vectors, batches by relation, and leverages closed-form orthogonal Procrustes updates for relational matrices—reducing training time and energy by orders of magnitude while maintaining competitive MRR/Hits@K performance (Peng et al., 2021).
• KGE-CL augments tensor factorization with contrastive learning to cluster semantically related entities and entity-relation pairs, boosting link prediction and embedding sparsity (Luo et al., 2021).

Robustness concerns include vulnerabilities to untargeted and targeted adversarial attacks (Zhao et al., 2024, Bhardwaj et al., 2021):

• Rule-based deletion (targeting structural motifs supporting confident rules) and adversarial addition (inserting negatives by corrupting low-confidence logical rules) degrade a wide spectrum of KGEs, with GNN- and rule-based methods exhibiting sensitivity contingent on graph density (Zhao et al., 2024).
• Poisoning via symmetry, inversion, or composition patterns can drastically reduce MRR/Hit@1 for models relying on these inductive biases (Bhardwaj et al., 2021).

6. Applications and Practical Ecosystems

KGE methods underpin a broad set of applications (Choudhary et al., 2021, Bianchi et al., 2020, Ge et al., 2023):

• Link prediction and KG completion: Standard MRR, Hits@K benchmarks on FB15k-237, WN18RR, YAGO3-10, ICEWS14/18, WikiArt-MKG.
• Recommender systems: Multi-modal, path-based, and attention-based KGEs (e.g., KGAT, MMGAT) improve user-item interaction modeling and collaborative filtering.
• Question answering and multi-hop reasoning: Embedding-based ranking of logical forms or stepwise question decomposition (EmbedKGQA, SimKGC).
• Entity clustering, classification, and query expansion: Embeddings supply semantic features for node-level tasks and document ranking (EDRM, Node Classification Benchmarks).

The KEEN Universe (Ali et al., 2020) offers reproducible infrastructure (PyKEEN, BioKEEN) with full experiment provenance, standardizes HPO, result reporting, and model sharing, and supports transfer to domains such as biomedicine.

Reliability estimation:

ReliK provides a black-box, task- and model-agnostic local reliability score per triple or subgraph—strongly correlated with downstream link prediction, classification, and rule mining performance, and computable efficiently via sampling (Egger et al., 2024).

7. Future Directions and Open Challenges

Outstanding research directions include:

• Unified algebraic/geometric foundations: Development of KGE models parameterized by structured operations (e.g., high-order Clifford algebras) that can capture complex logic and diverse relation properties (Cao et al., 2022, Xiao et al., 2024).
• Efficient hybridization of symbolic and neural inference: Models that tightly couple KGE with ontological rules, logical constraints, and dynamic or temporal updates (Niu, 2024, Cao et al., 2022).
• Continual and incremental multi-modal embedding: Scalable KGE adaptation to streaming, multi-modal KGs with non-uniform availability (Pons et al., 14 Nov 2025, Efthymiou et al., 2 Mar 2026).
• Explainability and fairness: Transparent post-hoc or model-intrinsic explanation tools, and reliable debiasing techniques (Bianchi et al., 2020).
• Integration with LLMs and structured prompting: Cross-model transfer to and from LLMs for zero-shot KGE and joint graph-language reasoning (Ge et al., 2023).
• Benchmarking, calibration, and reliability estimation: Systematic, locally valid evaluation protocols and measures such as Relik for subgraph-level analysis (Egger et al., 2024).

The trajectory of recent work signals a shift toward universal KGE architectures with an emphasis on semantic fidelity, multi-modal integration, efficient and robust optimization, and compatibility with evolving, heterogeneous, and explainable knowledge-centric systems.