ComplExEmbedding<T>
Models & Types · AiDotNet.RetrievalAugmentedGeneration.Graph.Embeddings
ComplEx embedding model: uses complex-valued embeddings with Hermitian dot product scoring.
Inheritance: KGEmbeddingBase<T> → ComplExEmbedding<T>
For Beginners
ComplEx is particularly good at modeling:
- Symmetric relations: “married_to” (score(A,r,B) = score(B,r,A))
- Antisymmetric relations: “parent_of” (score(A,r,B) ≠ score(B,r,A))
It achieves this by using complex numbers, which naturally distinguish direction.
How It Works
ComplEx (Trouillon et al., 2016) represents entities and relations as complex vectors. Score: Re(⟨h, r, conj(t)⟩) = Σ(hRe·rRe·tRe + hRe·rImag·tImag + hImag·rRe·tImag - hImag·rImag·tRe). Training uses logistic loss with optional N3 regularization. Higher scores indicate more plausible triples.
Complex numbers are represented as paired real/imaginary T[] arrays (length 2*dim each) to maintain generic T type parameter compatibility.
Constructors
| Constructor | Summary |
|---|---|
ComplExEmbedding() |
Properties
| Property | Summary |
|---|---|
bool IsDistanceBased { get; } |