DyHist: Query-Aware Historical Dynamics Modeling for Temporal Knowledge Graph Extrapolation

Problem setting

• TKG extrapolation infers a missing object in a future query by exploiting historical dependencies among entities and relations.The paper distinguishes extrapolation to future timestamps from interpolation over observed timestamps.Source: §1 · p. 1

Research gap

• Existing local-history methods often fuse earlier context and recent evidence into one evolving representation, which may obscure query-relevant shifts such as Oman appearing earlier and Djibouti more recently.The illustration uses Oman as medium-range context and Djibouti as more recent evidence near the query time.Source: §1 · p. 1, Figure 1 · p. 2
• Existing long-range methods may under-calibrate recurrence reliability: it can be useful for stable relations but less consistent for event-driven relations such as Accuse or Make statement.For event-driven relations, candidate entities may shift with recent events.Source: §1 · p. 2

Proposed approach

• DyHist combines query-aware recent and medium-range evidence with a relation-aware prior to model local changes and calibrate long-range recurrence.The prior uses time-decayed complete-fact recurrence, recent support trends, and learnable relation-specific strength.Source: §1 · p. 2

2.1 TKG Extrapolation

• TKG extrapolation methods include dynamic representation learning, temporal logical rules, generative approaches, temporal scoring, and textual semantics.Examples include Know-Evolve, RE-NET, RE-GCN, TLogic, DiffuTKG, NADEx, HTKGE, and AnRe.Source: §2.1 · pp. 2–3

2.2 Historical Information Modeling

• Historical-information methods span local evidence such as snapshots, neighborhoods, paths, and subgraphs, and long-range recurrence or global-context modeling.Local approaches can aggregate recent signals with earlier context, while recurrence methods may not calibrate reliability across relations and recent contexts.Source: §2.2 · p. 3
• DyHist distinguishes recent evidence, medium-range context, and long-range recurrence, using query-aware dual-window selection and residual injection locally and a relation-aware prior for recurrence.The prior calibrates recurring-pattern influence across relations and recent contexts.Source: §2.2 · p. 3

3.2 Historical Representation Encoding

• Local snapshots are encoded with RGCN structural representations and GRU transitions, while a sampled global subgraph supplies a query-time historical representation.The local entity-state matrix is updated over timestamps; the global representation is used for decoding.Source: §3.2 · p. 4

3.3 Query-aware Local Dynamics Modeling

• Query-aware selection chooses nonoverlapping recent and medium windows, aggregating near-query dynamics separately from earlier contextual evidence.The recent window is selected first, and the medium window is selected from the remaining local indices.Source: §3.3 · p. 4
• DyHist compares the latest short-term evidence with fused local context and injects their deviation through a learned gate as a local residual.The short-term evidence is obtained using the query-guided aggregator on the most recent subset.Source: §3.3 · p. 4

3.1 Overview

• DyHist combines local and global historical representations for decoding, then calibrates candidate logits with a relation-aware prior based on recurrence evidence.The prior uses time-decayed support, recent support changes, and learnable relation-specific recurrence strength.Source: §3.4 · pp. 4–5

3.4 Dynamic Historical Prior

• The historical prior calibrates candidate logits using time-decayed support, recent support changes, and relation-specific recurrence strength.A trend contribution is also controlled by ρ.Source: §3.4 · p. 5

3.5 Training Objective

• Training minimizes object-prediction negative log-likelihood and optimizes dual-window selection with prediction-oriented rewards, combining prediction and policy losses through λpg.The selection reward includes prediction-margin, loss-based, and temporal-role-separation feedback.Source: §3.5 · p. 5

4.1 Experimental Setups

• The evaluation uses chronological splits on ICEWS14, ICEWS05-15, ICEWS18, and GDELT, reporting MRR and Hits@{1, 3, 10} under time-aware filtering.Baseline comparisons include static, interpolation, and extrapolation methods.Source: §4.1 · p. 5

4.2 Main Results

• DyHist achieves the highest MRR on all four benchmarks, with competitive Hits@K results and the best MRR and Hits@10 on GDELT.Table 1 reports overall performance under the time-aware filtered setting.Source: Table 1 · p. 6, §4.2 · p. 6

4.3 Ablation Study

• Removing window selection and fast–slow residual injection, policy-gradient selection, or residual injection reduces performance; removing the prior causes the largest MRR drop among long-range variants.Removing time-decayed support or recent support trends also reduces performance.Source: §4.3 · p. 6

4.4 Analysis of Local History Modeling

• QADW achieves the best MRR on both datasets, while full query-aware descriptors outperform feature-shuffled or time-only variants.QADW adds about 0.08M parameters with similar inference time; descriptor analyses indicate query- and relation-specific correspondence provides additional guidance.Source: §4.4 · p. 7, Table 3 · p. 7, Table 4 · p. 7
• Adding fast–slow deviations improves on using the fused historical context alone, and gated residual injection yields a modest further benefit.Using only the most recent short-term representation performs worse than the dual-window context alone.Source: §4.4 · p. 7, Table 5 · p. 7

4.5 Repeated and Novel Queries

• Repeated queries outperform novel ones; DyHist improves over removing its prior on repeated queries and slightly outperforms Latest on novel queries.The results associate recurrence calibration with repeated queries and query-aware local evidence with novel queries.Source: §4.5 · pp. 7–8, Table 6 · p. 8

4.6 Analysis of Dynamic Historical Prior

• Learned relation-specific prior strengths perform best on ICEWS05-15, while a tuned fixed strength performs best on ICEWS14; learned strengths rise with coverage on ICEWS14 but not monotonically on ICEWS05-15.This indicates that coverage alone does not determine learned strength.Source: Figure 3 · p. 8, §4.6 · p. 8, Table 7 · p. 8

4.7 Computational Efficiency

• DyHist has the highest MRR on both efficiency-analysis datasets, with inference time close to LogCL on ICEWS14 and below CognTKE and MIDFA on ICEWS05-15.Its additional cost comes mainly from query-aware window selection and candidate-level prior calibration; the logit-level prior adds no deep decoding layers.Source: Figure 4 · p. 8, §4.7 · p. 8

• DyHist distinguishes query-relevant local changes from relation-dependent historical recurrence using adaptive dual windows, fast–slow residual modeling, and a relation-aware dynamic prior.The prior calibrates recurrence signals across relations and recent contexts.Source: §5 · pp. 8–9
• Experiments on four benchmarks support the usefulness of DyHist’s historical evidence modeling.Source: §5 · p. 9
• The results suggest that recurrence is useful, while query-aware local evidence remains important when direct recurrence is weak.Source: §5 · p. 9

• The four benchmarks mainly cover political and international events, leaving generalizability to other temporal knowledge graph domains unverified.The authors also note that recurrence and recent-support trends may be less effective for sparse or emerging events, and that recurrence calibration could amplify frequently observed entities or relations in sensitive decision-making.Source: §6 · p. 9

7 Ethical Considerations

• DyHist is intended solely for research on temporal reasoning and knowledge graph forecasting.Source: §7 · p. 9
• Because ICEWS and GDELT may have reporting biases and uneven event coverage, high-stakes use requires human oversight and domain-specific validation.Source: §7 · p. 9

A Notation Summary

• The notation summary covers symbols for snapshot modeling, query-aware window selection, local representation aggregation, and relation-aware prior calibration.Source: A Notation Summary · p. 11

Datasets and evaluation metrics

• Evaluation uses ICEWS14, ICEWS18, ICEWS05-15, and GDELT, with chronological splits and time-aware filtered MRR and Hits@{1, 3, 10}.Source: B Experimental Details · p. 11
• GDELT has 15-minute temporal granularity, while ICEWS05-15 has the most entities and timestamps among the ICEWS datasets.Source: B Experimental Details · p. 11

Additional implementation details

• Across datasets, the setup uses a local history length of 10, recent and medium windows of 3, and a fast window of 2.Source: B Experimental Details · p. 12

Training and inference

• Training uses timestamp-level batches, selects checkpoints by validation MRR, and uses top-ranked snapshots at inference for stable predictions.Source: B Experimental Details · p. 12

C Baseline Descriptions

• Baselines cover static knowledge graph reasoning, temporal interpolation, and temporal extrapolation methods.The descriptions include DistMult, ConvE, and RotatE; TTransE, TA-DistMult, and DE-SimplE; and a broad set of extrapolation models.Source: C Baseline Descriptions · p. 12

D Details of Query-aware Local Dynamics Modeling

• DyHist uses query-aware snapshot descriptors to select nonoverlapping recent and medium-range windows from local history.The supplementary section covers descriptor construction, dual-window selection, and policy optimization.Source: D Details of Query-aware Local… · p. 13, §D.1 · pp. 13–14

D.1 Snapshot Descriptor Construction

• Each snapshot descriptor combines query relevance, temporal lag, snapshot variation, latest-state consistency, and relation-context relevance.It also includes a recentness-adjusted relation-alignment feature.Source: §D.1 · pp. 13–14

D.2 Query-guided Temporal Aggregation

• Query-guided aggregation forms representations for the selected recent and medium windows, while a short-term representation provides fast evidence for residual deviation.Source: §D.2 · p. 14

D.3 Window Selection Loss

• Window selection is trained with a reward combining prediction feedback and temporal-role separation, then optimized using a policy-gradient objective.The margin reward assesses separation of the ground-truth entity from the strongest negative candidate; the NLL reward increases as prediction loss decreases.Source: §D.3 · pp. 14–15

E Repeated and Novel Queries

• A test query is repeated when its ground-truth object appeared with the same subject–relation pair in history; otherwise it is novel.The split is query-level and fixed across compared models.Source: E Repeated and Novel Queries · p. 15

F Details of Dynamic Historical Prior Calibration

• The historical prior retrieves facts sharing the query’s subject–relation pair, so only previously linked candidate objects receive historical support.Source: F Details of Dynamic Historical Prior… · p. 15

F.1 Query-conditioned Historical Support

• Time-decayed historical support gives more weight to recent occurrences while retaining long-range recurrence evidence.Source: §F.1 · p. 15

F.2 Relation-aware Prior Strength

• Relation-specific prior strength lets DyHist calibrate recurrence more strongly for repetitive relations and more weakly for event-driven relations.Source: §F.2 · p. 16

G Case Study

• The ICEWS14 case studies examine individual test queries and track ground-truth entity ranks across model variants or scoring stages using filtered ranks.Source: G Case Study · p. 16

G.1 Counterfactual Ranking Analysis

• The query is challenging because police-related and country entities receive high neural decoding scores.Source: §G.1 · p. 16
• Removing the dynamic historical prior lowers the ground-truth entity from rank 1 to 10, while DyHist ranks it first.The comparison evaluates variants on the same query in Table 11.Source: §G.1 · p. 16, Table 11 · p. 16
• Removing QADW replaces explicit window selection and fast-slow residual injection with query-aware aggregation over the whole local history, hurting the ground-truth rank.Source: §G.1 · p. 16

G.2 Detailed Evidence Utilization

• The detailed analysis examines selected historical evidence and prediction changes before and after relation-aware prior calibration.Source: §G.2 · p. 16
• The dynamic prior promotes the ground-truth entity from rank 19 to rank 1 in the successful-promotion case.Recent evidence captures statement-related local context, while medium evidence retrieves answer-related evidence involving Israeli Defense Forces.Source: §G.2 · p. 16, Table 12 · p. 17

G.3 Additional Case: Partial Promotion

• The partial-promotion case improves Iraq from rank 18 to rank 3 without making it the top prediction.Recent evidence contains Iraq-centered regional signals, and medium evidence includes an earlier consultation-related event involving Iraq.Source: §G.3 · p. 17, Table 13 · p. 17

G.4 Discussion

• Recent and medium windows provide complementary evidence, and the dynamic prior can promote supported objects without acting as a hard copy rule.Across the cases, recent windows reflect near-query contexts while medium windows provide earlier semantic clues.Source: §G.4 · pp. 17–18