Traditional knowledge graphs (KGs) represent the world as a collection of static facts, structured as triples: $(\text{subject}, \text{relation}, \text{object})$. For example, a graph might contain the triple $(\text{Barack Obama}, \text{PresidentOf}, \text{USA})$. While mathematically clean, this formulation suffers from a fundamental flaw: the real world is dynamic. Relationships emerge, change, and expire. Treating temporal facts as universal, timeless truths prevents models from distinguishing between what was true and what is true, leading to obsolete or incorrect predictions in rapidly shifting domains like finance, news, and medicine.
To solve this, researchers have moved from static vector spaces to temporal knowledge representations. Instead of merely mapping entities and relations to fixed coordinates, these models incorporate time directly into the geometry of the embedding space.
Slicing Time into Hyperplanes
One intuitive way to handle time-varying facts is to treat time as a geometric modifier. In HyTE: Hyperplane-based Temporally aware Knowledge Graph Embedding, Shib Sankar Dasgupta et al. propose dividing the timeline into discrete intervals, with each interval associated with its own distinct hyperplane.
Why does this work? In a standard translation-based embedding model, the goal is to satisfy the algebraic relation $\mathbf{h} + \mathbf{r} \approx \mathbf{t}$, where $\mathbf{h}$, $\mathbf{r}$, and $\mathbf{t}$ are vectors for the head, relation, and tail. If an entity possesses different relationships over time, a single static vector cannot satisfy all of them without collapsing the distinct meanings of those relations.
By projecting the static entity and relation embeddings onto a time-specific hyperplane $P_t$, the model dynamically alters their coordinates based on when the interaction occurs. The optimization objective transforms into:
$$\mathbf{h}_{\perp t} + \mathbf{r}_{\perp t} \approx \mathbf{t}_{\perp t}$$
This projection allows the same entity to exhibit different relational behaviors in different eras. For instance, the vector for a politician will align closely with a specific political office during their term's hyperplane, but drift away in hyperplanes representing later years. This effectively disambiguates relations that are semantically similar but temporally distinct.
Continuous Evolution via Point Processes
While discrete time-slicing works well for coarse intervals, real-world events occur continuously. Slicing time too finely creates data sparsity, while slicing too coarsely blurs the exact sequence of events.
To capture the fluid, event-driven nature of reality, Rakshit Trivedi et al. introduced Know-Evolve: Deep Temporal Reasoning for Dynamic Knowledge Graphs. Instead of static snapshots, the authors model the occurrence of facts as a multi-variate point process.
The core intuition is that the formation of a relationship is an event that depends on the historical trajectory of the entities involved. Know-Evolve uses a conditional intensity function to modulate the probability of an edge forming between two entities at a precise timestamp $t$. When an event occurs, the embeddings of the participating entities are instantly updated using a non-linear recurrent mechanism.
By learning how entity representations evolve over continuous time, the model doesn't just predict whether a relationship exists; it models the temporal dynamics well enough to predict exactly when a future interaction is likely to happen.
Enforcing Temporal Logic
Purely geometric or probabilistic models can still make logically impossible predictions—such as predicting that an individual holds a political office before they are born. To ground these embeddings in reality, models must incorporate structural and chronological constraints.
In Towards Time-Aware Knowledge Graph Completion, Tianyu Liu et al. demonstrate that combining temporal embeddings with an explicit optimization step drastically improves link prediction. The authors use an Integer Linear Programming (ILP) framework to enforce logical temporal constraints during inference. By ensuring that predicted facts respect chronological order and mutual exclusivity (e.g., a person cannot hold two mutually exclusive offices simultaneously), the model filters out geometrically plausible but logically impossible links, reducing prediction error significantly compared to static baselines.
From Static Tensors to Episodic Memory
Another powerful paradigm for knowledge graph completion is tensor factorization, where a knowledge graph is represented as a three-dimensional adjacency tensor. Models like TuckER, introduced by Ivana Balazevic et al. in TuckER: Tensor Factorization for Knowledge Graph Completion, use Tucker decomposition to learn expressive, linear representations of static graphs.
To capture dynamic facts within this algebraic framework, Erik A. Daxberger et al. developed temporal tensor models, such as "ConT", in Embedding models for episodic knowledge graphs. By introducing a fourth dimension—time—into the tensor decomposition, these models learn to factorize episodic knowledge. This mathematical formulation provides a computational parallel to cognitive science: it demonstrates how stable semantic memory (timeless concepts) can naturally emerge from the aggregation of dynamic episodic memories (time-stamped events).