Source-linked AI summary
BridgeMem: Causal Dyadic Transition Residuals for Temporal Knowledge Graph Forecasting
Zeyan Li, Libing Chen, Shengda Zhuo, Yin Tang, Jianfeng Xu
TL;DR
Temporal knowledge graph forecasters can miss how prior relations on a specific actor–candidate pair change the odds of a target relation. BridgeMem adds a support-adaptive dyadic transition residual to a frozen forecaster. Across five benchmarks, it improves on the strongest prior result in all 20 filtered metric comparisons.
Problem
Existing forecasters often summarize entity states, paths, or exact recurrence without representing how prior relations on the query actor–candidate pair shift the target-relation odds.
Source: §1 · p. 1Method
BridgeMem retrieves strictly prior ordered-pair events, combines neural and count-based transition estimates by support, and adds an uncertainty-gated likelihood-ratio residual to a frozen forecaster.
Source: §1 · p. 2, §2.1 · p. 2, §2.2 · p. 2Results
Across five benchmarks, BridgeMem improves MRR over the strongest prior result by 0.0213 on ICEWS 2014 and wins all 20 filtered MRR and Hits@{1, 3, 10} comparisons.
Source: §3.2 · p. 3Takeaways & Limitations
BridgeMem provides dyadic transition evidence complementary to, rather than a replacement for, the backbone forecaster’s entity-level evidence.
Source: §4 · p. 4The method uses a fixed K = 8 event window and τ = 100 support threshold across datasets, with the gate and λ selected on validation rather than learned.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Temporal knowledge graph forecasting aims to infer future relational facts from the temporal structure of observed events. Existing forecasters mainly summarize history through entity states, relation states, paths, or exact recurrence. These views often miss pair-specific transition evidence, that is, the way prior relations between the query actor and a candidate change the odds of the target relation. We introduce BridgeMem, which estimates this quantity as a residual added to the log scores of a frozen full-vocabulary forecaster. For each candidate, BridgeMem retrieves the pair's events that strictly precede t, encodes their relations, directions, and lags, and converts them into a likelihood-ratio correction. A support-adaptive empirical-Bayes reader trusts exact transition counts where they are abundant and backs off to a learned attention estimator where they are sparse. The backbone's own uncertainty gates the correction, so confident queries and candidates without dyadic history are left unchanged. On five benchmarks, BridgeMem improves on the strongest of nine baselines from 2021--2026 in all 20 filtered MRR and Hits@{1,3,10} comparisons, with MRR gains of 0.0213, 0.0164, 0.0216, 0.0112, and 0.0028 over the best prior result. These results show the value of explicit dyadic transition modeling.
1. INTRODUCTION
Temporal knowledge graph forecasters rank candidates from past events, but entity-state and recurrence views can miss how relations on a specific actor–candidate pair shift the odds of a target relation. BridgeMem models that dyadic transition evidence as a candidate-specific residual on a frozen forecaster.
1. INTRODUCTION
- Entity-state and recurrence views do not explicitly represent how prior relations on the ordered actor–candidate pair shift the odds of a target relation.Entity states omit which events involved the query actor, while copy distributions require the exact target relation to have occurred on the pair.Source: §1 · p. 1
- A useful dyadic memory must preserve transition sequences while distinguishing reliable repeated transitions from sparse one-off observations.Exact lookup alone risks frequency noise.Source: §1 · p. 1
- BridgeMem estimates candidate-specific dyadic transition evidence as a likelihood-ratio residual on a frozen full-vocabulary forecaster.It retrieves strictly prior pair events and combines relation, direction, and recency evidence using count and attention estimates.Source: §1 · p. 2
- Across five benchmarks, BridgeMem improves over nine prior methods on all filtered MRR and Hits@{1, 3, 10} comparisons.The residual is added without end-to-end retraining.Source: §1 · p. 2
2. BRIDGEMEM
BridgeMem retrieves causal histories for ordered actor–candidate pairs, estimates their transition evidence with neural and count-based readers, and adds a support-adaptive residual to a frozen forecaster. The correction is gated by the backbone’s uncertainty and is trained and selected without retraining the backbone.
2.1. Causal Dyadic State
- The causal dyadic state stores up to eight recent events between the ordered actor–candidate pair within a 65-snapshot window, with relation, orientation, and lag information.A strict boundary excludes events at the forecast time; candidates without qualifying history keep their backbone score.Source: §2.1 · p. 2
2.2. Neural and Count-Based Transition Evidence
- The neural reader pools dyadic events, while smoothed counts estimate relation transitions conditioned on direction and recency relative to the marginal relation prior.The support-dependent mixture trusts empirical counts when support is high and backs off to the learned estimate when support is sparse.Source: §2.2 · p. 2
2.3. Uncertainty-Gated Residual
- The uncertainty gate scales the likelihood-ratio residual, vanishing for confident oracle decisions and peaking when oracle uncertainty is 0.5.The backbone remains fixed, and candidates without dyadic history are unchanged.Source: §2.3 · p. 2
2.4. Training and Selection
- The transition encoder is trained self-supervised from strictly earlier dyad events, and validation selects its epoch, residual settings, and gate while keeping the backbone frozen.Inference requires at most K indexed reads per candidate.Source: §2.4 · pp. 2–3
3. EXPERIMENTS
The experiments compare BRIDGEMEM with nine prior methods across five filtered temporal knowledge graph benchmarks. BRIDGEMEM improves every reported metric, and removing its transition residual reduces performance while the residual adds limited fitting cost.
3.1. Experimental Setup
- The evaluation covers ICEWS 2014, ICEWS 2018, ICEWS 2005–2015, GDELT, and WIKI, using filtered query-micro MRR and Hits@{1, 3, 10} on complete query sets.The nine rerun baselines share the processed files and evaluation protocol; the matched backbone is the released CENET model and checkpoint for each dataset.Source: §3.1 · p. 3
3.2. Main Results
- Improvements concentrate at H@1 and H@3 rather than low-ranked entities; GDELT gains were 0.0121/0.0108/0.0103 at H@1/H@3/H@10 over more than 610k bidirectional test queries.WIKI was near saturation under CENET, but all four metrics still improved, with most remaining gain at H@1.Source: §3.2 · p. 3
- MRR gains over the strongest prior result were 0.0213 on ICEWS 2014, 0.0164 on ICEWS 2018, 0.0216 on ICEWS 2005–2015, 0.0112 on GDELT, and 0.0028 on WIKI; all 20 filtered comparisons improved.The final table block reports BRIDGEMEM and its per-metric improvement over the strongest prior method.Source: §3.2 · pp. 3–4, Table 1 · p. 4
3.3. Mechanism Control and Efficiency
- Removing the learned and count-based residual while keeping the backbone, candidates, filter, and queries fixed lowers all 20 metrics; effects are strongest on ICEWS 2005–2015 and ICEWS 2014, and remain positive on GDELT.Figure 3 presents test-set gains from the complete residual over identical frozen CENET scores.Source: Figure 3 · p. 4, §3.3 · p. 4
- On WIKI, with 24 relations and backbone MRR 0.8681, dyadic evidence raises H@1 by 0.0047 without reducing broader cutoffs.Source: §3.3 · p. 4
- The memory adds 16.7k–91.4k parameters, and runtime scales with sampled transitions rather than entity vocabulary; GDELT’s final train-plus-validation refit took 657 seconds.The GDELT selection plus train-only refit took 1,424 seconds; other final refits finished in under six minutes.Source: §3.3 · p. 4, Table 2 · p. 4
4. CONCLUSION
BridgeMem adds pair-specific transition evidence as a residual to a frozen forecaster, complementing entity-level and path-level scoring. It improves all reported comparisons across five benchmarks, with gains largest on datasets with rich multi-year interaction histories.
4. CONCLUSION
- All 20 reported MRR and Hits@{1, 3, 10} comparisons improve over the strongest of nine prior methods across five benchmarks.The residual adds 16.7k–91.4k parameters.Source: §4 · p. 4
- Gains are largest on datasets with rich multi-year interaction histories and remain positive on WIKI despite the backbone being near saturation.Source: §4 · p. 4
- The dyadic transition signal is complementary to what entity-level and path-level forecasters already capture.Source: §4 · p. 4
- The method fixes the event window at K = 8 and support threshold τ = 100 across datasets, with the gate and λ selected on validation rather than tuned per dataset.Source: §4 · p. 4