Temporal Knowledge Graphs: Facts That Change Over Time
Published:
From Triples to Quadruples
Standard KG: {(s, r, o)} — timeless facts.
Temporal KG: {(s, r, o, t)} where t is a timestamp or interval [t_start, t_end].
Examples:
- (Barack_Obama, presidentOf, USA, [2009, 2017])
- (Bayern_Munich, wonChampionsLeague, 2020)
- (Apple, ceoIs, Steve_Jobs, [1976, 1985] ∪ [1997, 2011])
Two types of TKG facts:
- Instantaneous: single timestamp (sports results, news events)
- Interval-based: valid during a period (job titles, relationships)
The TKG Completion Task
Interpolation: predict missing facts at known historical times — fill in KG gaps within the training period.
Extrapolation: predict future facts — given everything known up to time t, what triples will be true at t+1?
Extrapolation is the harder and more practically relevant task.
Key Models
TTransE (Time-aware TransE)
Adds time to the TransE scoring function:
Learns a separate time embedding w_t and treats time as another “relation” that shifts entity positions. Simple extension but ignores temporal dynamics.
TNTComplEx
Extends ComplEx to quadruples by adding temporal embeddings as additional mode:
(4th-order tensor decomposition with complex embeddings.)
RE-NET (Recurrent Event Network)
Uses a recurrent architecture to model temporal sequences:
- For each entity pair (s, o), collect the sequence of relations over time
- Encode with GRU/LSTM to get history representation
- Aggregate with a GNN over the event subgraph at time t
- Score candidate triples for t+1
RE-NET explicitly models the temporal ordering of events — capturing recurrence patterns like “Player X scores goals in consecutive matches.”
TGAT (Temporal Graph Attention Network)
Assigns time-encoding to edges and applies attention over temporal neighbourhoods. Each neighbour message is weighted by both structural importance (attention) and temporal proximity (time encoding).
Worked Example: TTransE on a Political Event
Suppose we want to predict: (CountryX, holdsElection, ?, t=2024).
Static TransE embeds CountryX and holdsElection without time — it either always predicts elections or never does, based on training frequency.
| TTransE scoring function: f(s, r, o, t) = - | e_s + w_r + w_t - e_o |
With learned time embedding w_2024 ≈ w_2020 (both election years, similar temporal position in the 4-year cycle), TTransE will score 2024 candidates similarly to 2020 — capturing the periodicity pattern. For a non-election year like 2022, w_2022 is far from w_2020 in embedding space, so election predictions score low.
This illustrates why time embeddings help: they encode position in recurring cycles, giving the model a “calendar sense” that static embeddings completely lack.
Temporal Reasoning Challenges
1. Irregular observation: facts are not observed at uniform time intervals — some entities have dense histories, others sparse.
2. Time granularity: a fact valid for decades appears at daily/monthly resolution differently than a single-day event.
3. Entity dynamics: entities change identity over time (companies merge, people change roles). The embedding of “CEO of Apple” should change as different people hold the role.
4. Causality vs correlation: temporal patterns in KGs often reflect causal chains, but models learn correlations. Disentangling these is an open problem.
TKG Benchmarks
- ICEWS (Integrated Crisis Early Warning System): political events worldwide, timestamped daily
- GDELT: global event database, fine-grained temporal resolution
- YAGO15K: static YAGO with temporal annotations
- WikiData (temporal subset): entity facts with validity intervals
Standard splits: train on t ≤ T, validate on T < t ≤ T’, test on t > T’.
Summary
| Model | Approach | Temporal pattern captured |
|---|---|---|
| TTransE | Time embedding + TransE | Basic time displacement |
| TNTComplEx | 4-order tensor with time | Complex temporal interactions |
| RE-NET | GNN + RNN | Temporal event sequences |
| TGAT | Temporal attention | Recency-weighted neighbourhood |
Temporal knowledge graphs are a stepping stone from static relational reasoning to full temporal graph learning (covered in the Dynamic Graphs section). The key insight: facts have lifetimes, and reasoning about the world requires reasoning about when facts were true — not just whether they are true.
References
- Lacroix, T., Obozinski, G., & Usunier, N. (2020). Tensor Decompositions for Temporal Knowledge Base Completion. ICLR 2020 (TNTComplEx).
- Jin, W., Qu, M., Jin, X., & Ren, X. (2020). Recurrent Event Network: Autoregressive Structure Inference over Temporal Knowledge Graphs. EMNLP 2020 (RE-NET).
- Xu, D., Ruan, C., Körpeoglu, E., Kumar, S., & Achan, K. (2020). Inductive Representation Learning on Temporal Graphs. ICLR 2020 (TGAT).