Source-linked AI summary
Temporal Knowledge Graph Completion: A Survey
Borui Cai, Yong Xiang, Longxiang Gao, He Zhang, Yunfeng Li, Jianxin Li
TL;DR
Knowledge graphs are incomplete, and static KGC cannot adequately represent facts that change over time. This paper surveys TKGC by organizing methods around timestamp integration and covering its background, benchmarks, evaluation, limitations, and future directions. The survey identifies timestamp-aware modeling as a basis for improved link prediction while highlighting unresolved scalability, sampling, data, and continual-learning challenges.
Problem
Knowledge graphs are widely incomplete, while static KGC cannot adequately capture facts whose truth changes over time.
Method
The paper provides the first survey of TKGC, covering its definition, datasets, evaluation metrics, timestamp-based methods, limitations, and future directions.
Results
TKGC methods have shown improved link prediction accuracy over KGC methods on multiple knowledge graph datasets.
Takeaways & Limitations
The survey organizes TKGC research by how timestamps are integrated to capture temporal dynamics for link prediction.
Takeaways & Limitations
TKGC remains constrained by challenging time-aware negative sampling, benchmark-data limitations, computational cost at real-life scale, and evolving knowledge graphs.
Abstract
from arXiv · showhide
Knowledge graph completion (KGC) can predict missing links and is crucial for real-world knowledge graphs, which widely suffer from incompleteness. KGC methods assume a knowledge graph is static, but that may lead to inaccurate prediction results because many facts in the knowledge graphs change over time. Recently, emerging methods have shown improved predictive results by further incorporating the timestamps of facts; namely, temporal knowledge graph completion (TKGC). With this temporal information, TKGC methods can learn the dynamic evolution of the knowledge graph that KGC methods fail to capture. In this paper, for the first time, we summarize the recent advances in TKGC research. First, we detail the background of TKGC, including the problem definition, benchmark datasets, and evaluation metrics. Then, we summarize existing TKGC methods based on how timestamps of facts are used to capture the temporal dynamics. Finally, we conclude the paper and present future research directions of TKGC.
1 Introduction
Knowledge graphs support many applications but are widely incomplete, while conventional KGC assumes facts are static even though facts can change over time. TKGC incorporates timestamps to capture temporal dynamics, and this paper surveys the task, methods, datasets, metrics, and future directions.
- Knowledge graphs represent real-world events as structured facts and support recommender systems, question answering, and information retrieval.
- Existing knowledge graphs are widely incomplete, with more than 70% of Freebase person entities missing a known place of birth.
- KGC infers missing facts from existing ones, but its static-fact assumption limits its ability to learn temporal dynamics.
- TKGC incorporates fact timestamps to improve link prediction by capturing temporal dynamics that static KGC cannot represent.
- The paper surveys TKGC background, benchmark datasets, evaluation protocols, timestamp-based methods, conclusions, and future research directions.
2 Background
TKGC represents timestamped facts and evaluates missing entities, relations, or timestamps by ranking candidates with factual scores. Its background includes benchmark datasets, time-aware filtering, and ranking losses used to separate true facts from negatives.
- A timestamped knowledge graph contains entities, relations, timestamps, and facts represented as quadruples {h, r, t, τ}.
- TKGC commonly treats timestamps as discrete time points, although facts may originally use time intervals.
- Margin ranking, cross-entropy, and binary cross-entropy losses separate true facts from negative samples in different ways.
- Benchmark Datasets: ICEWS and GDELT use time points, whereas WIKIDATA and YAGO15K use time intervals; YAGO15K is harder because many facts lack timestamps.
- Evaluation ranks the true candidate using factual scores and commonly reports Hits@k, Mean Ranking, and Mean Reciprocal Ranking.
- Entity prediction evaluates missing heads and tails, while relation and timestamp prediction extend the task to missing relations or temporal information.
3 Temporal Knowledge Graph Completion Methods
TKGC extends KGC scoring models because timestamped facts require representations that capture temporal information in addition to entity and relation structure.
- TKGC models score quadruples by learning temporal representations alongside entity and relation representations.
3.1 Timestamp-included Tensor Decomposition
Timestamp-included tensor decomposition adds time as a fourth tensor mode, enabling TKGC models to factorize interactions among entities, relations, and timestamps. CP and Tucker variants provide different representation flexibility, while multivector models support point and interval timestamps.
- Adding timestamps converts the knowledge graph from a 3-way entity-relation tensor into a 4-way tensor.
- Canonical Polyadic Decomposition: CP decomposition scores a fact with the four-way interaction q(s) = ⟨eh, er, et, eτ⟩ and can use an imaginary timestamp for static facts.
- Canonical Polyadic Decomposition: Multivector CP representations model timestamp intervals through separate start- and end-time relation representations.
- Tucker Decomposition: Tucker decomposition factorizes a tensor through a core tensor and mode-specific matrices, generalizing CP decomposition.
- Tucker Decomposition: Tucker TKGC scores interactions among entities, relations, and timestamps while relaxing the requirement that their embedding dimensions be equal.
3.2 Timestamp-based Transformation
Timestamp-based transformation methods adapt entity and relation representations to specific times rather than treating them as static. They encode time through synthetic relations, geometric transformations, or sequence-based representations.
- Timestamp-based transformations learn entity and relation representations corresponding to specific times, addressing changing contexts in temporal knowledge graphs.
- Synthetic relations concatenate timestamps with relations, converting temporal quadruples into triples that existing KGC models can process.For example, {Lakers, championOf, NBA, 2010} becomes {Lakers, championOf:2010, NBA}.
- Optimal timestamp construction can split intervals at detected change points or merge intervals when related facts remain consistent.This reduces unnecessary synthetic relations when facts have overlapping time intervals.
- Explicit timestamp tokens can be encoded with an LSTM to produce more expressive synthetic relation representations.The relation and timestamp are represented as a token sequence such as {championOf, 2y, 0y, 1y, 0y}.
- Hyperplane projections map static entity or relation representations into time-specific regions, while relational matrices and GRUs can add relation specificity and temporal dynamics.The hyperplane projection is applied before computing factual scores, and GRUs process sequences of hyperplanes.
- Complex-space rotations transform entity representations at a timestamp before evaluating the relation as a translation between the rotated entities.This approach uses Hermitian products and supports time-dependent factual scoring.
3.3 Dynamic Embedding
Dynamic embedding methods model temporal knowledge graph representations as evolving over time rather than as independent time-specific states. They use explicit temporal patterns or recurrent processes to capture how entities and relations change.
- Dynamic embeddings aim to capture structured evolution patterns because entity and relation representations across time are interdependent.The paper illustrates this with an irreversible life-cycle sequence such as bornIn → workAt → dieIn.
- Time-function models decompose representations into static, trend, seasonal, and noise components.A hyperbolic-space variant combines static and time-varying components, interpreting evolution as movement on manifolds.
- Recurrent neural networks can adaptively learn dynamic evolution instead of imposing a specific temporal pattern.They are used to represent the occurrence of facts as a multidimensional temporal point process.
- A conditional intensity function scores facts to infer the next occurrence time, while separate RNNs update the head and tail entity representations.Each RNN receives inputs incorporating previously observed fact information.
3.4 Learning from Knowledge Graph Snapshots
Snapshot-based methods represent a temporal knowledge graph as a sequence of timestamped subgraphs and learn how graph states evolve across that sequence. Their models use Markov transitions, autoregressive generation, and graph-based neighborhood information.
- Timestamps partition a knowledge graph into snapshots, turning link prediction into inference over changing multi-relational interactions.Each snapshot contains facts associated with its corresponding timestamp.
- A first-order Markov model makes each snapshot state depend on the previous state through a probability transition matrix.The model factors the joint probability over graph states, entities, and relations across time.
- The Markov-based model learns expressive evolution patterns that are effective for TKGC in both offline and online scenarios.
- Autoregressive methods generate facts in a snapshot from preceding snapshots and recurrently incorporate local multi-hop neighbors with a multi-relational graph aggregator.The autoregressive order determines how many preceding snapshots provide graph-structure context.
3.5 Reasoning with Historical Context
Historical-context methods use the chronological order of timestamped facts to reason about missing links. They differ in how they identify which prior facts are relevant, using attention, temporal displacement, or domain-informed tendencies.
- Historical context consists of earlier facts related to a query and provides the basis for temporal reasoning in missing-link prediction.
- Attention-based reasoning expands a query-dependent inference subgraph by sampling neighboring historical facts and weighting them with edge attention scores.The expansion is directed toward the query’s interests.
- Temporal-displacement reasoning captures both whether a fact is past, present, or future and the exact magnitude of that displacement.A two-stage GNN incorporates these temporal features with structural information from query-related entities and relations.
- Domain-informed methods organize historical facts using Goodness and Closeness scores that encode relation hostility and entity cooperation.Goodness distinguishes relations such as sanction and collaborate, while Closeness measures cooperation between entities.
4 Conclusion and Future Directions
The paper identifies several directions for improving TKGC, including external knowledge, time-aware negative sampling, scalability, and continual learning for evolving graphs. These directions address limitations in current datasets, training procedures, computational scale, and static evaluation settings.
- 4.1 Incorporate External Knowledge: External knowledge can enrich limited structural and temporal information, especially for long-tailed entities and relations.Relational domain knowledge, entity types, and pretrained language models are proposed as sources of additional information.
- 4.2 Time-aware Negative Sampling: Time-aware negative sampling remains underexplored, despite its importance for effective representation learning in TKGC.The additional time dimension makes it necessary to model interactions between facts and timestamps when generating negative samples.
- 4.3 Larger-scale Knowledge Graph: TKGC must become more efficient and scalable because real-world knowledge graphs can contain billions of facts, unlike current benchmark datasets.The paper highlights distributed computation and compositional embeddings as promising directions for reducing training and parameter costs.
- 4.4 Evolutionary Knowledge Graph: Continual learning is needed because real-world knowledge graphs evolve through fact deletion and inclusion rather than remaining invariant.Experience replay and knowledge distillation have been explored for catastrophic forgetting, while other continual-learning techniques remain open directions.