Source-linked AI summary
C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning
Yassir Lairgi, Ludovic Moncla, Khalid Benabdeslem, Rémy Cazabet, Pierre Cléau
TL;DR
Existing weak-signal detectors omit semantic, relational, or temporal structure, and dynamic TKG reasoning has not addressed emergence detection. C-Unseen uses LLM reasoning to extract narrative-deviating rare subgraphs and track their persistence across snapshots. Experiments report better performance with self-interpretable outputs, including on an anonymized benchmark variant.
Problem
Existing keyword, topic, graph, and TKG methods omit semantic or relational context, use untyped topology, or focus on link prediction rather than weak-signal detection.
Method
C-Unseen uses an LLM Rare Subgraphs Extractor and a Weak Signal Alerter to identify narrative-deviating rare subgraphs and track their corroboration across time.
Results
C-Unseen achieves better performance with self-interpretable outputs, and retains performance on a named-entity-anonymized benchmark variant.
Takeaways & Limitations
The framework provides a self-interpretable approach for detecting weak signals in Dynamic Temporal Knowledge Graphs.
Takeaways & Limitations
The benchmark supplements Wikipedia’s incomplete pre-signal history with 16 manually added atomic facts.
Abstract
from arXiv · showhide
Weak signals are early, low-visibility indicators that precede significant changes before those changes become established. Existing detection methods, based on keyword frequency, topic modeling, or untyped graph topology, fail to capture the semantic and relational structure through which such signals manifest. In this paper, we propose C-Unseen, a self-interpretable framework for weak signal detection in Dynamic Temporal Knowledge Graphs (DTKGs). We define a weak signal as a rare, semantically coherent subgraph that proliferates across consecutive TKG snapshots. The framework operates through two modules: a Rare Subgraphs Extractor, in which an LLM identifies subgraphs whose content is in tension with the dominant snapshot narrative via chain-of-thought reasoning, and a Weak Signal Alerter, in which the persistence of these rare subgraphs is tracked across time steps to isolate true weak signals. Experimental results demonstrate that C-Unseen outperforms keyword-, topic-, and graph-based baselines.
1 Introduction
Weak signals are fragmentary precursors whose significance emerges through corroboration, but existing keyword, topic, and topology methods miss their semantic and relational structure. C-Unseen addresses this gap with LLM-based rare-subgraph extraction and temporal persistence tracking.
- Weak signals are fragmentary, low-visibility precursors whose significance emerges only through corroboration.
- Keyword and topic methods reduce domain knowledge to statistical distributions, while topological methods fail to capture connections between entities.
- Dynamic Temporal Knowledge Graphs preserve timestamped entity-relation triples, retaining semantic and temporal structure for changing sources.
- C-Unseen extracts rare subgraphs in tension with each snapshot’s dominant narrative using LLM chain-of-thought reasoning.
- The Weak Signal Alerter tracks rare subgraphs across time steps to isolate persistent weak signals.
2 Related works
Prior work represents weak signals as terms, topics, or structures, but each representation omits part of the semantic, relational, or temporal information needed for dynamic TKG detection.
- Keyword methods identify weak signals through low visibility and high growth across successive time windows.
- Topic-modeling methods represent weak signals as small thematic clusters and track their evolution over time.
- Graph-based methods detect peripheral or anomalous structures using co-occurrence topology, spanning trees, cliques, and graphlets.
- TKG reasoning methods target link prediction, while sparse-interaction aggregation can dilute rare-entity representations.
- Existing approaches respectively lose context, relational structure, typed information, or a mechanism for weak-signal detection.
3 Proposed method
C-Unseen models dynamic temporal knowledge as evolving snapshots of typed quintuples, identifies rare subgraphs against snapshot narratives, and tests their temporal corroboration through connected context.
- Definitions: A DTKG is an ordered sequence of TKG snapshots observed at timestamps, with typed entities, relations, and time-bounded quintuples.
- Definitions: The model distinguishes observation time from each fact’s inherent validity period.
- Definitions: A rare subgraph contains quintuples whose content conflicts with the snapshot’s dominant narrative.
- Definitions: A connecting subgraph joins entities in a rare subgraph through the union of their shortest entity-level paths.
- Definitions: A weak signal combines snapshot-level rarity with corroboration from a related rare subgraph at an earlier observation time.
- Rare Subgraphs Extractor: Module 1 prompts an LLM to summarize each snapshot, compare quintuples with that narrative, and identify rare subgraphs.
- Weak Signal Alerter: Module 2 compares the current connecting subgraph with prior connecting subgraphs to determine which deviations persist as weak signals.
4 Experiments
The experiments evaluate C-Unseen’s detection accuracy, lead time, interpretability, and component contributions on the Wiki-OpenAI benchmark. C-Unseen provides broader signal coverage, self-interpretable explanations, and performance gains associated with rare-subgraph selection and chain-of-thought prompting.
- Experimental design: The evaluation measures detection accuracy, lead time, interpretability, and component contributions through four research questions.Detection accuracy uses shared-anchor-word thresholds k ∈ {1, 2, 3}; lead time measures early-warning value relative to strong-signal establishment.
- Dataset: The Wiki-OpenAI benchmark contains 773 atomic facts from Wikipedia and manually added news-derived facts, with five annotated strong signals.Wikipedia facts cover monthly edits from 2015–2025; duplicates are removed and 16 facts are added to address missing pre-signal history.
- RQ2: Lead Time: At k = 3, C-Unseen’s mean lead time is 1.00 year versus 1.70 ± 0.45 years for BERTrend.The comparison uses 4.20 ± 1.10 covered signals for C-Unseen and 1.40 ± 0.55 for BERTrend, so the means reflect different coverage levels.
- RQ3: Interpretability: C-Unseen represents the weak signal as a TKG subgraph with a natural-language interpretation tracing its evolution across snapshots.The explanation exposes structural tensions, including a transition from an abstract safety boundary to tension between policy and defence interests.
- RQ4: Ablation: Selecting rare subgraphs and their connecting subgraph yields ∆F1 = +0.141±0.093 over the full-DTKG arm, while full-DTKG input adds no measurable gain.The ablation attributes the gain to the combination of selected rare subgraphs and the CoT prompt rather than the LLM alone.
5 Conclusion
C-Unseen is a self-interpretable framework for weak signal detection in Dynamic Temporal Knowledge Graphs, with DTKGs serving as temporal knowledge storage, persistent reasoning memory, and structural context. Experiments on Wiki-OpenAI report better performance and self-interpretable outputs, retained on a named entity anonymized benchmark variant, while scalability to large DTKGs remains unresolved.
- C-Unseen formally defines weak signals in Dynamic Temporal Knowledge Graphs and provides a self-interpretable detection framework.
- DTKGs store domain knowledge over time, preserve intermediate LLM reasoning across snapshots, and connect rare subgraphs for the Weak Signal Alerter.
- Experiments on Wiki-OpenAI show better performance with self-interpretable outputs, retained on a named entity anonymized benchmark variant.
- Scaling the Rare Subgraphs Extractor to large DTKGs remains an open issue because full snapshots may exceed the LLM context window.