Knowledge Graph Embedding (KGE) models aim to represent entities and relations in a continuous vector space to predict missing links and reason over structured knowledge. Early paradigms treated relations as simple spatial translations: given a fact triplet (head, relation, tail), the model enforced the geometric intuition that the head vector plus the relation vector should approximate the tail vector ($h + r \approx t$).
While translation is highly intuitive, it suffers from a fundamental mathematical limitation: it struggles to preserve complex logical patterns. For instance, if a relation is symmetric (such as is_spouse_of), the translation framework requires $h + r = t$ and $t + r = h$, forcing the relation vector $r$ to collapse to zero and mapping both distinct entities to the exact same point. To overcome these geometric bottlenecks, researchers have shifted toward modeling relations as geometric transformations—such as rotations, reflections, and affine maps—across complex, hypercomplex, and non-Euclidean spaces.
Rotations in Complex and Hypercomplex Spaces
The core intuition behind rotation-based embeddings is that logical properties can be naturally mapped to algebraic invariants. In a 2D complex plane, multiplying a vector by a complex number of unit length rotates that vector by a specific angle. This operation elegantly resolves the symmetry dilemma: a rotation of $\pi$ radians is its own inverse. Thus, a relation can map a head entity to a distinct tail entity while perfectly satisfying symmetry without collapsing the underlying vector space.
Sun et al. formalized this in RotatE by representing relations as element-wise rotations in a complex vector space via the Hadamard product. This geometric formulation allows a single model to simultaneously satisfy symmetry, antisymmetry, inversion (via a reverse rotation), and composition (where the addition of two rotation angles yields a third). This rotation paradigm has proven highly effective across diverse domains, from modeling intricate biomedical interactions in frameworks like CoV-KGE (Zeng et al.) and the COVID-19 knowledge graph developed by Huang et al., to neighborhood homogenization in graph neural networks like TransGCN (Cai et al.).
Translation (Euclidean): [Head] ------ +Relation ------> [Tail] (Collapses if symmetric)
Rotation (Complex): [Head] \
\ rotated by θ
\----------> [Tail] (Preserves distinct vectors)
However, standard complex numbers are limited to a two-dimensional plane, meaning their multiplications are commutative ($A \cdot B = B \cdot A$). Real-world relations are frequently non-commutative; for example, the composition of mother's brother does not equal brother's mother. To capture these multi-faceted, non-commutative interactions, Zhang et al. introduced QuatE, which extends entity and relation representations into a 4D hypercomplex quaternion space. By utilizing the Hamilton product, the model gains the expressive power to handle non-commutative compositions with fewer parameters.
Building on this hypercomplex foundation, Wang et al. developed DaBR, which applies bidirectional rotations to maximize entity separability. To scale these transformations beyond fixed dimensions, Shimodaira et al. proposed OrthogonalE, modeling relations as block-diagonal orthogonal matrices. This generalizes low-dimensional rotations to arbitrary dimensionalities while keeping the underlying transformations computationally manageable.
Beyond Rotation: Reflections, Projections, and Rules
Rotations belong to a broader class of distance-preserving transformations known as isometries. Other isometric operations offer distinct advantages for relational differentiation. For instance, Mao et al. proposed RREA, which leverages orthogonal relational reflections. Because reflections are inherently orthogonal, they allow the model to flip entity distributions across a hyperplane, achieving relational differentiation without distorting the underlying data distribution.
While rotations easily handle symmetry and inversion, they struggle with transitivity (if $A \rightarrow B$ and $B \rightarrow C$, then $A \rightarrow C$). In a pure rotation space, adding angles indefinitely does not naturally enforce a shortcut constraint. To bridge this gap, Song designed Rot-Pro, which couples relational rotations with idempotent projections. By projecting entity vectors into a relation-specific subspace where the transitive property holds, the framework successfully captures all five major logical patterns—symmetry, asymmetry, inversion, composition, and transitivity—simultaneously.
This geometric structure can also be used to explicitly bind symbolic logic rules to continuous spaces. In RulE, Tang et al. embedded first-order logical rules by constraining the combined rotation of antecedent relations to match the rotation of the consequent relation. This algebraic constraint enables robust, noise-resilient "soft" logical reasoning. Similarly, He et al. introduced PConE, which models relations as rotations confined within a cone-based geometric space, providing an inductive bias that allows models to reason through multi-hop logical queries more effectively.
Scaling and Affine Transformations for Density and Boundaries
Isometries preserve distances, but real-world knowledge graphs often feature high-density connectivity, such as many-to-one or one-to-many relations (e.g., thousands of distinct individuals born in the same city). Mapping a wide array of distinct head entities to a single tail entity using pure distance-preserving transformations can cause severe spatial crowding and degradation.
To handle high-density connectivity, models must incorporate scaling operations. Li et al. addressed this in STaR by combining scaling with translation and rotation, allowing the model to shrink or expand entity regions depending on the relation's density. Yang et al. similarly demonstrated that applying relation-specific affine transformations (scaling and translation) to entity vectors allows simpler translational models to satisfy complex geometric requirements.
Affine Transformation: [Entity Point/Box] ---> [Scale & Shift] ---> [Transformed Boundary]
When entities are represented as spatial regions rather than single points, affine transformations become essential for manipulating boundaries. In BEUrRE, Chen et al. modeled entities as $n$-dimensional hyperrectangles ("boxes"). Relations are treated as affine transformations that scale and shift these boxes, allowing the model to calculate calibrated probabilistic volumes to reason over uncertain knowledge graphs. Fang et al. extended this box-affine paradigm to temporal knowledge graphs in PTBox, using polynomial approximations to dynamically scale and shift entity boundaries across continuous time intervals.
Escaping Flat Space: Hyperbolic and Conformal Manifolds
Knowledge graphs frequently exhibit hierarchical structures, such as taxonomies or ontologies (e.g., canary $\rightarrow$ bird $\rightarrow$ animal). In flat Euclidean space, the volume of a sphere grows polynomially with its radius, whereas the number of nodes in a hierarchical tree grows exponentially with its depth. This mathematical mismatch causes severe distortion and crowding when embedding deep hierarchies in flat space.
Hyperbolic space natively resolves this issue because its volume expands exponentially relative to its radius, providing "more room" at its fringes to accommodate tree-like structures. To capture both hierarchical structures and logical patterns, Chami et al. introduced ATTH, which combines hyperbolic geometry with relational rotations, reflections, and attention mechanisms.
Early hyperbolic models frequently mapped operations back and forth through Euclidean tangent spaces, introducing numerical instability and geometric distortion. To bypass these pitfalls, Fan et al. developed LorentzKG and Liang et al. developed FHRE. Both frameworks define rotations and reflections directly on the Lorentz manifold, preserving radial distances and capturing complex graph patterns entirely within the hyperbolic space. To extend this to multi-entity interactions (such as $n$-ary relations), Yan et al. introduced PolygonE, which models multi-entity facts as "gyro-polygons" within the Poincaré ball using a combination of rotations and Möbius additions.
Euclidean Space: Volume grows polynomially --> Crowds hierarchical trees
Hyperbolic Space: Volume grows exponentially --> Fits hierarchical trees naturally
To resolve relational multiplicity—where multiple distinct relations exist between the same pair of entities—models must combine different types of spatial movements. Shimodaira et al. developed 3H-TH, which integrates 3D quaternion rotations with Möbius addition-based translations in hyperbolic space. Similarly, Zheng et al. introduced HolmE, combining rotation and Möbius addition to ensure that the relation space remains mathematically closed under composition.
Finally, Zhang et al. pushed this geometric evolution to the Riemann sphere with 5EL. By leveraging conformal Möbius group transformations, 5EL unifies translation, rotation, scaling, inversion, and reflection within a single projective framework, allowing temporal knowledge graphs to model evolving logical patterns over time.
Key Foundational Work
RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space (2019) — contributed 2 concept elements to this direction.
Repurpose Open Data to Discover Therapeutics for COVID-19 using Deep Learning (2020) — contributed 1 concept elements to this direction.
Quaternion Knowledge Graph Embeddings. (2019) — contributed 1 concept elements to this direction.