In a traditional knowledge graph, a fact is a triplet: (Subject, Predicate, Object). For example, (Shohei Ohtani, playsFor, Los Angeles Angels). But the real world is not static. Facts have lifespans. Ohtani played for the Angels until 2023, then moved to the Dodgers.
To capture this, Temporal Knowledge Graphs (TKGs) add a fourth dimension: time. However, simply adding a timestamp is not enough for a machine to understand change. Recent research has shifted from treating entities as static points in vector space to treating them as dynamic trajectories that evolve over time.
From Static Points to Temporal Functions
The core intuition behind this research direction is that an entity’s identity is composed of two parts: a permanent core and a time-varying state. Instead of representing an entity as a fixed vector $\mathbf{e}$, researchers like Rishab Goel and Mehran Kazemi in Diachronic Embedding for Temporal Knowledge Graph Completion propose representing entities as functions of time, $\mathbf{e}(t)$.
By using sinusoidal functions (sine and cosine), these models can capture both linear trends (a player getting older) and periodic patterns (seasonal sports cycles). This approach, also explored by Benyou Wang in Sequential Modeling in Vector Space, allows the model to "calculate" the state of an entity at any arbitrary point in time, even if no fact was recorded at that exact moment.
Geometry of Evolution
If an entity is moving through latent space, how do we describe that movement? Researchers have proposed several geometric mechanisms to "evolve" an embedding:
- Rotations and Quaternions: In RotateQVS, Kai Chen and colleagues represent time as a rotation in a 4D quaternion space. As time passes, the entity's vector rotates. This is mathematically elegant because it preserves the "norm" (the magnitude) of the entity while changing its orientation relative to other entities.
- Velocity Vectors: Zhen Han and Volker Tresp in DyERNIE model entity evolution as movement on Riemannian manifolds. They assign each entity a "velocity vector" in a tangent space, allowing entities to follow complex paths—like hierarchies or cycles—more naturally than they could in flat Euclidean space.
- Hyperplane Projections: Models like RTGE, discussed by Yonghui Xu in Time-Aware Graph Embedding, project static embeddings onto time-specific hyperplanes. This effectively "slices" the latent space by time, ensuring that a fact like "Barack Obama is President" is only valid within the geometric slice corresponding to 2009–2017.
Disentangling History and Dynamics
A major challenge in predicting future facts (extrapolation) is distinguishing between what an entity is and what it is doing.
Caren Han and Kunze Wang address this in Re-Temp by decomposing embeddings into static properties, linear trends, and periodic seasonal functions. Similarly, the HIP Network by Peng Zhang disentangles structural interactions from repetitive behaviors.
This disentanglement is crucial for tasks like KG-FRUS, a dataset of 127 years of US diplomacy. By tracking the "geopolitical drift" of nations as dynamic embeddings, researchers can quantify how alliances shift from colonial structures to regional independence movements over a century.
Continuous Time via Latent ODEs
While many models treat time as discrete steps (e.g., days or years), the real world is continuous. Shandian Zhe and Mike Kirby pushed this boundary in THIS-ODE by using Latent Ordinary Differential Equations (ODEs).
Instead of jumping from state to state, the model learns a neural network that represents the derivative of the entity's state. To find an entity's position at time $t$, the model integrates this derivative. This allows for "time alignment," where interactions occurring at irregular intervals (like a user visiting a store) can be modeled with high precision, capturing the underlying logic of the interaction rather than just the sequence.
Why This Works
The shift toward dynamic trajectories works because it respects the temporal smoothness of the real world. Entities rarely change their nature overnight; their roles and relationships evolve along a path. By modeling the path (the trajectory) rather than the points (the snapshots), these models can:
- Interpolate: Guess what happened between two known timestamps.
- Extrapolate: Predict where an entity's trajectory is headed in the future.
- Reason: Answer complex questions like "Who was the leader before the revolution?" by navigating the temporal geometry of the graph, as shown in TempoQR by Costas Mavromatis.