ComplEx

ComplEx operates in complex vector space (ℂⁿ), where each embedding has real and imaginary components. This allows the model's scoring function to be non-symmetric, a critical property for modeling diverse relation types. The core scoring function is defined as the real part of a trilinear dot product: score = Re(⟨h, r, conj(t)⟩), where h, r, t are complex vectors for head, relation, and tail, and conj(t) is the complex conjugate of the tail embedding. This formulation is the mathematical key to its expressiveness.

A primary advantage of ComplEx is its inherent ability to model antisymmetric relations. In a real vector space, a symmetric scoring function (where score(h,r,t) = score(t,r,h)) is common, which fails for relations like isParentOf. Because ComplEx uses the complex conjugate of the tail entity, the scores for (h, r, t) and (t, r, h) are generally not equal, allowing it to correctly learn that isParentOf(Alice, Bob) does not imply isParentOf(Bob, Alice). This directly addresses a major limitation of earlier models like DistMult.

Despite its strength with antisymmetric patterns, ComplEx can also effectively model symmetric relations (e.g., isMarriedTo, collaboratesWith) and inverse relations (e.g., isChildOf is the inverse of isParentOf).

• Symmetry emerges when the relation embedding is purely real, causing the scoring function to become symmetric.
• Inversion is captured when one relation's embedding is approximately the complex conjugate of another. This multi-pattern capacity within a single, unified model simplifies architecture and improves generalization across diverse knowledge graphs like FB15k and WN18.

The model employs an efficient bilinear scoring mechanism. While the embedding space is complex, the scoring function can be computed as a special case of a real bilinear product, enabling efficient implementation. The decomposition Re(⟨h, r, conj(t)⟩) expands into operations on the real and imaginary parts of the vectors, allowing for optimized tensor operations using standard deep learning libraries like PyTorch and TensorFlow. This efficiency enables training on large-scale graphs with millions of entities and relations.

ComplEx is a semantic matching model. It predicts links by measuring the compatibility between the head, relation, and tail embeddings in the latent space, rather than by simulating a path or translation. To predict a missing tail for a query (h, r, ?), the model scores the head and relation against all candidate entity embeddings, ranking them by the computed score. This approach is evaluated using standard metrics like Hits@k and Mean Reciprocal Rank (MRR), where ComplEx established state-of-the-art results on benchmark datasets upon its introduction.

ComplEx's complex space formulation inspired a lineage of more advanced models. RotatE (2019) reinterpreted relations as rotations in complex space (t = h ◦ r, where |r| = 1), explicitly modeling symmetry/antisymmetry, inversion, and composition. QuatE (2019) extended the idea into hypercomplex (quaternion) space for richer representational capacity. ComplEx's core insight—that moving beyond real-valued embeddings unlocks critical relational patterns—established a fundamental direction for subsequent knowledge graph embedding research, influencing both academic and industrial KGC systems.

ComplEx excels at predicting asymmetric relations, where the relationship direction matters. This is critical for business knowledge graphs.

• Example: Predicting (Apple, competesWith, Samsung) is true, while (Samsung, competesWith, Apple) is also true (symmetric). However, (Apple, acquires, Beats) is true, but (Beats, acquires, Apple) is false (asymmetric). ComplEx's complex dot product naturally captures this directional asymmetry.
• Use Case: In a supply chain knowledge graph, correctly distinguishing (Supplier, providesPartTo, Manufacturer) from its inverse is essential for dependency analysis and risk modeling.

The model effectively learns inverse relations, where one relation logically reverses another (e.g., hypernym/hyponym, parentOf/childOf).

• Mechanism: If a relation r and its inverse r_inv exist, ComplEx can learn embeddings where the score for (h, r, t) is high and the score for (t, r_inv, h) is also high, as the complex conjugate provides a natural mathematical inverse.
• Application: In pharmaceutical knowledge graphs, this ensures consistency between (Drug, treats, Disease) and (Disease, treatedBy, Drug), improving the completeness of therapeutic relationship data.

ComplEx embeddings power entity-aware recommendation engines by representing users, items, and their interactions (e.g., likes, purchases) as a knowledge graph.

• Process: Users and items are entities. Relations like clicked, purchased, or rated_high connect them. ComplEx predicts missing purchased links, suggesting new items.
• Advantage: Its handling of asymmetric relations improves scenarios where userA influences userB does not imply userB influences userA. This is vital for modeling social influence in retail and content platforms.

In molecular knowledge graphs, ComplEx predicts interactions between biological entities, a task rich with complex relation patterns.

• Data Structure: Entities are genes, proteins, diseases, and drugs. Relations include interactsWith, inhibits, associatedWith.
• Prediction Task: Given a known protein and a disease, predict the targets relation to a novel drug candidate. ComplEx's ability to reason over multi-hop paths (e.g., Gene → associatedWith → Disease ← treatedBy → Drug) aids in identifying novel therapeutic pathways.

ComplEx is deployed to auto-populate corporate knowledge graphs, inferring missing relationships between internal entities like projects, employees, skills, and departments.

• Example: Inferring that (Project_Alpha, requiresSkill, MachineLearning) based on similar project profiles, or predicting (Employee_B, reportsTo, Manager_Y) from organizational patterns.
• Business Value: This automates the maintenance of HR, IT, and R&D knowledge bases, ensuring data completeness for downstream applications like expert finding or team composition analysis.

ComplEx is a staple baseline and competitive solution on public KGC benchmarks, demonstrating its empirical effectiveness.

• Common Datasets: WN18RR (WordNet), FB15k-237 (Freebase). These datasets contain mixtures of symmetric, asymmetric, and inverse relations.
• Typical Metrics: ComplEx consistently achieves strong Hits@10 (proportion of correct answers in top 10 predictions) and Mean Reciprocal Rank (MRR) scores on these benchmarks, validating its design for real-world relation patterns. Its efficiency also makes it suitable for large-scale enterprise graphs.

Knowledge Graph Embedding (KGE) is the foundational technique of mapping entities and relations into a continuous vector space. This transformation enables mathematical operations for tasks like link prediction. ComplEx is a specific KGE model that uses complex-valued vectors to capture a richer set of relation patterns (symmetric, asymmetric, inverse) compared to real-valued models.

• Purpose: Create a latent representation where geometric relationships between embeddings correspond to logical relationships in the graph.
• Core Idea: If a triple (h, r, t) is true, the embedding of the head entity h should be close to the embedding of the tail entity t after some transformation defined by the relation r.
• Applications: Powering downstream tasks like semantic search, recommendation systems, and data completion.

Link prediction is the primary task that ComplEx is designed to solve. It involves predicting the existence of a missing relationship between two entities in a knowledge graph. For a given query (head, relation, ?) or (?, relation, tail), the model scores candidate entities to identify the most plausible completion.

• Formal Task: Given (h, r, ?), predict the correct tail entity t. Given (?, r, t), predict the correct head entity h.
• ComplEx's Role: ComplEx provides a scoring function φ(h, r, t) that measures the plausibility of a triple. A higher score indicates a more likely true fact.
• Example: In a biomedical KG, link prediction can infer new gene-disease associations not present in the original data.

TransE and RotatE are other seminal KGE models that, like ComplEx, define specific geometric transformations for relations.

• TransE: Models relations as translations in the vector space: h + r ≈ t. It is simple and effective but struggles with complex relation patterns like 1-to-N.
• RotatE: Models relations as rotations in complex vector space: t ≈ h ◦ r, where ◦ denotes element-wise (Hadamard) multiplication. It can model symmetry, anti-symmetry, and inversion.
• ComplEx vs. RotatE: While both use complex numbers, RotatE constrains relation embeddings to have unit modulus (a rotation), whereas ComplEx does not, allowing it to also model composition patterns effectively.

DistMult and RESCAL are bilinear models that form the direct lineage of ComplEx. ComplEx was developed to overcome a key limitation of DistMult.

• DistMult: Uses a simple, efficient diagonal matrix to represent each relation. Its scoring function is a trilinear dot product: h · diag(r) · t. A major flaw: it can only model symmetric relations because the dot product is commutative.
• RESCAL: A more expressive predecessor that uses a full relation matrix, capturing all pairwise interactions but requiring many more parameters (O(d^2)).
• ComplEx as a Solution: By moving to complex space and using the Hermitian dot product, ComplEx regains the ability to model asymmetric relations (like DistMult cannot) while remaining parameter-efficient (like DistMult, O(d)).

Tensor factorization is the overarching mathematical framework for many KGE models, including ComplEx. A knowledge graph can be represented as a 3D binary tensor X of size (n_entities, n_relations, n_entities), where a slice for each relation r is an adjacency matrix.

• The Goal: Factorize this sparse, high-dimensional tensor into lower-dimensional, dense embeddings that capture its latent structure.
• ComplEx Factorization: ComplEx performs a canonical polyadic (CP) decomposition of the tensor, where each element X[h,r,t] is approximated by the real part of the sum of the products of the embedding components: Re(∑ h_i * r_i * conj(t_i)).
• Advantage: This formulation directly leads to its efficient scoring function and its theoretical ability to model key relation patterns.

Negative sampling is a critical training technique used to optimize ComplEx and other KGE models. Since knowledge graphs contain only positive examples (true triples), the model needs contrasting negative examples to learn meaningful distinctions.

• Process: For each true triple (h, r, t) in a training batch, one or more corrupted triples are generated by randomly replacing the head or tail entity with another entity from the graph (e.g., (h', r, t) or (h, r, t')).
• Loss Function: ComplEx is typically trained with a margin-based ranking loss or binary cross-entropy loss. The objective is to score true triples higher than corrupted ones by at least a defined margin.
• Challenge: Generating informative, "hard" negatives (those that are plausible but false) is an active research area to improve model performance.