Knowledge Graph Embedding

The foundational mechanism for many embedding models, treating relationships as vector translations in the latent space.

• TransE: Interprets a relation r as a translation from head entity h to tail entity t, optimizing for h + r ≈ t
• TransR: Projects entities into a relation-specific space before applying the translation, handling complex relational patterns
• RotatE: Models relations as rotations in complex vector space, capturing symmetry, antisymmetry, and inversion patterns

These geometric constraints force the model to learn structurally consistent representations that generalize to unseen triples.

A class of models that represent knowledge graphs as multi-dimensional tensors and decompose them to uncover latent semantics.

• RESCAL: Factorizes the adjacency tensor into entity and relation matrices, capturing pairwise interactions between all entities for each relation
• DistMult: A simplified bilinear diagonal model that restricts relation matrices to be diagonal, reducing parameters while maintaining expressiveness for symmetric relations
• ComplEx: Extends DistMult into complex-valued space, enabling the modeling of asymmetric and antisymmetric relations through the use of complex conjugate operations
• TuckER: Applies Tucker decomposition to the full tensor, learning a core tensor that encodes multi-relational interactions

Tensor methods excel at capturing global statistical patterns across the entire graph.

Modern approaches that use message-passing neural networks to generate embeddings by aggregating information from neighboring nodes.

• R-GCN: Extends Graph Convolutional Networks to handle multiple relation types, applying relation-specific weight matrices during neighborhood aggregation
• CompGCN: Jointly learns entity and relation embeddings through composition operations before convolution, reducing parameter explosion in highly multi-relational graphs
• Node Classification: GNN-based embeddings propagate label information through the graph structure, enabling zero-shot classification of new entities based on their relational context
• Inductive Capability: Unlike transductive methods, GNN encoders can generate embeddings for unseen entities at inference time by leveraging their local neighborhood structure

This inductive property is critical for cold-start scenarios where new items continuously enter the catalog.

The critical training technique that teaches embeddings to distinguish true facts from plausible falsehoods by corrupting valid triples.

• Uniform Negative Sampling: Randomly replaces the head or tail entity with any other entity in the graph, providing a simple but noisy training signal
• Bernoulli Sampling: Adjusts the corruption probability based on relation cardinality, reducing false negatives for one-to-many relationships
• Adversarial Sampling: Dynamically selects hard negative examples that the current model scores highly, forcing the embeddings to learn finer-grained distinctions
• Self-Adversarial Sampling: Weights negative samples by their current model score during loss computation, avoiding the computational cost of explicit hard negative generation

The quality of negative samples directly determines the discriminative power of the resulting embeddings.

The primary downstream task that leverages learned embeddings to infer missing relationships between entities, directly addressing the item cold-start problem.

• Tail Prediction: Given a new item entity and a relation type, the model ranks all candidate tail entities to find the most likely connections, such as new_product → belongs_to → category
• Head Prediction: Identifies which existing entities should link to a new item, enabling reverse relationship inference
• Triple Classification: Binary verification of whether a proposed fact (h, r, t) is true, used to validate automatically extracted knowledge
• Zero-Shot Transfer: Embeddings trained on a rich product taxonomy can immediately position a new SKU within the semantic space based solely on its categorical and attribute relationships

This bypasses the need for any user interaction history by exploiting the dense web of entity relationships.

The technique of identifying equivalent entities across disparate knowledge graphs, enabling cross-domain cold-start recommendations by merging semantic spaces.

• Embedding Space Alignment: Learns a linear or non-linear transformation that maps embeddings from one graph onto another using a small set of seed alignments
• Iterative Bootstrapping: Alternates between aligning entities with high confidence and retraining the mapping, progressively expanding the set of matched pairs
• Multi-Modal Fusion: Incorporates entity attributes, names, and descriptions alongside structural embeddings to improve alignment precision
• Cross-Lingual Transfer: Aligns knowledge graphs in different languages, allowing a product catalog in one market to bootstrap recommendations in a new geographic region

This enables a retailer expanding internationally to immediately position new products using knowledge transferred from established markets.