Paper deep dive
C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning
Yassir Lairgi, Ludovic Moncla, Khalid Benabdeslem, Rémy Cazabet, Pierre Cléau
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/28/2026, 4:22:30 AM
Summary
The paper introduces C-Unseen, a framework for detecting weak signals in Dynamic Temporal Knowledge Graphs (DTKGs) using Large Language Models (LLMs). It defines weak signals as rare, semantically coherent subgraphs that persist across time. The framework uses two modules: a Rare Subgraphs Extractor to identify subgraphs in tension with the dominant narrative via Chain-of-Thought reasoning, and a Weak Signal Alerter to track persistence across snapshots. It outperforms keyword, topic, and graph-based baselines on a new benchmark dataset, Wiki-OpenAI.
Entities (11)
Relation Signals (8)
C-Unseen → processes → Dynamic Temporal Knowledge Graphs
confidence 95% · framework for weak signal detection in Dynamic Temporal Knowledge Graphs (DTKGs).
C-Unseen → usesmodule → Weak Signal Alerter
confidence 95% · The framework operates through two modules: ... and a Weak Signal Alerter...
C-Unseen → usesmodule → Rare Subgraphs Extractor
confidence 95% · The framework operates through two modules: a Rare Subgraphs Extractor...
Weak Signal → definedas → Rare Subgraph
confidence 90% · We define a weak signal as a rare, semantically coherent subgraph that proliferates across consecutive TKG snapshots.
Wiki-OpenAI → usedforevaluation → C-Unseen
confidence 90% · We introduce what is, to the best of our knowledge, the first benchmark for weak signal detection... The dataset, referred to as Wiki-OpenAI...
Rare Subgraphs Extractor → usestechnique → Chain-of-Thought
confidence 90% · in which an LLM identifies subgraphs ... via chain-of-thought reasoning
C-Unseen → outperforms → BERTrend
confidence 85% · Experimental results demonstrate that C-Unseen outperforms keyword-, topic-, and graph-based baselines.
C-Unseen → outperforms → BEAM
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2608.26870v1
- Canonical: https://arxiv.org/abs/2608.26870v1
Trouble viewing inline? Open PDF directly →
Full Text
40,932 characters extracted from source content.
Expand or collapse full text
C-Unseen: Weak Signal Detection in Dynamic Temporal Knowledge Graphs via LLM Reasoning Yassir LAIRGI✉ Affiliation: INSA Lyon, CNRS, UCBL, LIRIS, UMR5205, 69621 Villeurbanne, France E-mail ludovic.moncla, khalid.benabdeslem, remy.cazabet@liris.cnrs.fr Affiliation: GAUC, Lyon, France E-mail yassir.lairgi, pierre.cleau@auvalie.com Ludovic MONCLA Affiliation: INSA Lyon, CNRS, UCBL, LIRIS, UMR5205, 69621 Villeurbanne, France E-mail ludovic.moncla, khalid.benabdeslem, remy.cazabet@liris.cnrs.fr Khalid BENABDESLEM Affiliation: INSA Lyon, CNRS, UCBL, LIRIS, UMR5205, 69621 Villeurbanne, France E-mail ludovic.moncla, khalid.benabdeslem, remy.cazabet@liris.cnrs.fr Rémy CAZABET Affiliation: INSA Lyon, CNRS, UCBL, LIRIS, UMR5205, 69621 Villeurbanne, France E-mail ludovic.moncla, khalid.benabdeslem, remy.cazabet@liris.cnrs.fr Pierre CLÉAU Affiliation: GAUC, Lyon, France E-mail yassir.lairgi, pierre.cleau@auvalie.com Abstract 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-Unseen11 1 The code, prompts, and dataset are available at https://github.com/AuvaLab/itext2kg, 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. Keywords: Weak Signal Detection Dynamic Temporal Knowledge Graphs Large Language Models. 1 Introduction In strategic domains, decisions must often be made before clear evidence emerges. The concept of weak signals addresses this challenge [4]. A strong signal is unambiguous evidence that a change has already taken place: its effects are visible, and its consequences can be assessed with confidence. A weak signal is its precursor, appearing while the change is only beginning to take shape and too fragmentary on its own to prove that anything is happening at all. The example in Figure 1 illustrates this distinction. Weak signals are characterized across three measurable dimensions: visibility, degree of emergence, and interpretability [12]. Their significance, however, does not exist in isolation: it arises from their connections to other entities and events within a broader context [3]. Example: from a rare element to a strong signal Rare (10 January 2024). OpenAI quietly removes “military and warfare” from the prohibited-uses list in its usage policy. No announcement. Reported by a single outlet. Corroboration and potential weak signal (16 January 2024). At Davos, OpenAI confirms work with DARPA on cybersecurity tools, and states the blanket ban had blocked legitimate use cases. Strong signal (16 June 2025). A US Department of Defense contract (CDAO), reportedly $200M, alongside the launch of “OpenAI for Government”. The policy edit was individually dismissible; its significance lay in corroboration Figure 1: Weak signals precede a strong signal as fragmentary, low-visibility traces whose significance emerges only from their corroboration. Existing methods are primarily keyword frequency-based. They evaluate the distribution of terms across documents to classify them as weak, latent, or strong signals [26]. More recent approaches leverage transformer-based topic models within online learning frameworks to classify topics along a noise-to-signal continuum [6, 10]. However, these methods reduce rich domain knowledge to statistical distributions over topics, failing to capture connections between entities. To incorporate structural dynamics, topological approaches have been proposed. For instance, [2] introduced a framework that uses "graphlets", small, localized subgraph patterns within temporal interaction networks, as structural signatures to detect weak signals. While this approach effectively models the temporal evolution of network interactions, it operates on simple, untyped graphs. Consequently, it still lacks a semantic layer; it cannot represent named entities, typed relations, or the semantic context of how knowledge evolves. Temporal Knowledge Graphs (TKGs) offer a natural alternative. A TKG encodes a set of documents as timestamped entity-relation triples, for example (WHO, issued alert on, monkeypox, 2022-05-20), preserving the semantic and temporal layer that the previous methods discard. A TKG, however, is built from the documents available at one moment, whereas monitoring weak signals means reading a source that keeps changing. This raises two challenges. First, many facts assert no validity period at all, so there is no date to attach to them other than the date they were observed. Second, without a dynamic update mechanism, it would be difficult to track events that appear and disappear over time. A dynamic TKG (DTKG) [14] addresses both by using dual-time modeling, separating when a fact is observed from its validity period. For instance, an article from January 23, 2020 reporting that a virus had spread to ten countries is stored with the observation time tobs=t_obs= 2020-01-23 and an empty validity period, since the article states no period of its own and the fact may have been true weeks earlier. Because weak signals appear at a given date and gain substance only as later information corroborates them, a DTKG offers a solid foundation for detecting them. Despite this potential, applying DTKGs to weak signal detection faces a practical limitation: constructing DTKGs from unstructured text has traditionally required domain-specific pipelines and extensive human annotation [27]. Recent zero- and few-shot LLM-based methods have substantially reduced this cost [15, 20, 14]. However, no existing method addresses the next step: using a dynamic TKG to detect weak signals. TKG reasoning methods are systematically biased toward frequent patterns and focus on link prediction rather than emergence detection [7, 25]. In this paper, we propose a framework for weak signal detection in dynamic TKGs. Our contributions are: (i) to the best of our knowledge, this is the first definition of weak signals in dynamic TKGs; (i) we introduce a self-interpretable framework that operates through two distinct modules: a Rare Subgraphs Extractor, in which each TKG snapshot is processed by an LLM via chain-of-thought (CoT) reasoning to identify subgraphs whose content is in tension with the dominant snapshot narrative, and a Weak Signal Alerter, in which these rare subgraphs are tracked across time steps to isolate weak signals. In the remainder of the paper, we present related work in Section 2, the proposed approach in Section 3, experimental evaluation in Section 4, and a conclusion in Section 5. 2 Related works Early approaches to weak signal detection typically begin by extracting candidate keywords using TF-IDF weighting. These candidates are then evaluated according to two complementary criteria: their occurrence frequency and their temporal growth across successive time windows. Specifically, keywords are considered weak signals when they exhibit both low occurrence frequency, captured by the Degree of Visibility (DoV), and a high growth rate, captured by the Degree of Diffusion (DoD) [26]. The resulting weak signals are subsequently projected onto a BCG-style matrix to support expert interpretation. This framework has been replicated across diverse domains [21, 16]. However, these keyword-based methods lose the context that gives each term its meaning. To address the lack of context limitation of isolated terms, many methods propose topic modeling under the hypothesis that a weak signal manifests as a small cluster that is disjoint from dominant topics. LDA constitutes the predominant technique [19], extended to the temporal dimension via dynamic topic modeling and topic chain construction, wherein similarity measures track thematic evolution across successive time windows [22, 11]. More recently, transformer-based topic models such as BERTopic and BERTrend have been proposed to monitor shifts within online learning architectures, continuously classifying topics along a noise-to-signal continuum [6]. Despite their richer context, all these approaches operate at topic-level granularity, remaining blind to the relational structure that may exist between the entities of the same topic. Early graph-based methods construct keyword co-occurrence graphs and apply betweenness, degree, and proximity to isolate terms that occupy structurally peripheral yet topologically influential positions as candidate precursors [13, 9]. Complementary approaches leverage minimum spanning trees to reduce graph complexity while preserving signal-bearing connectivity patterns [13], and employ 3-cliques and density metrics to characterize early structural anomalies in domain-specific interaction networks [5]. Moreover, [1] proposes BEAM, a framework that enumerates graphlets, all connected induced subgraphs of two to five nodes, across temporally ordered graph snapshots, and characterizes weak signals through the velocity and acceleration of graphlet count series combined with a rarity ratio. While these methods represent a significant advance in structural weak signal detection, they operate on homogeneous, untyped graphs and do not exploit the semantic layer of typed entities and relations. Current TKG reasoning methods addressing link prediction, such as RE-GCN [17], utilize recurrent graph convolutions; however, their neighborhood aggregation mechanisms tend to dilute sparse interactions, resulting in poor representations of rare entities. While DPCL-Diff [8] addresses historical sparsity through node diffusion, it assumes new events resemble existing ones, which limits its ability to identify truly novel patterns. LLM-based frameworks like LLM-DA [23] and GenTKG [18] have introduced rule-guided forecasting, yet they remain biased toward frequent patterns and lack dedicated rarity scoring mechanisms for weak signals. These approaches are optimized for link prediction, leaving them limited in detecting the rare aspect of weak signals. Despite these advances, current methods face the following limitations: (1) keyword-based approaches capture the weak signal as a term while losing the context that gave each term its meaning. (2) Topic modeling approaches recover that context at the theme level while remaining insensitive to relational structure. (3) Graph-based methods exploit topology but operate on homogeneous, untyped graphs. (4) TKG reasoning frameworks are optimized for link prediction, providing no mechanism to detect weak signals. To address these challenges, we propose a framework for weak signal detection in dynamic TKGs. 3 Proposed method 3.1 Definitions Dynamic Temporal Knowledge Graph. Following [14], a Dynamic Temporal Knowledge Graph (DTKG) is defined as a sequence of TKG snapshots stt∈obs\G_s^t\_t _obs, where obsT_obs is an ordered set of observation timestamps. Each snapshot is defined as: st=(ℰt,ℛt,start,endt,ℱt)G_s^t= (E^t,\,R^t,\,T_start^t,\,T_end^t,\,F^t ) where ℰtE^t is the set of typed entities, ℛtR^t the set of typed relations, and ℱtF^t the set of quintuples of the form (es,r,eo,tstart,tend)(e_s,r,e_o,t_start,t_end), with es,eo∈ℰte_s,e_o ^t, r∈ℛtr ^t, and tstart∈start_start _start^t, tend∈endtt_end _end^t denoting the validity period of the fact. The dual-time modeling distinguishes between the observation time t and the inherent validity period of each fact. Rare Subgraph. A rare subgraph at observation time t is defined as a set of quintuples t⊆ℱtS^t ^t whose content is in tension with the dominant narrative of stG_s^t. Connecting Subgraph. Let ℰ(t)E(S^t) denote the set of entities appearing in the rare subgraph tS^t. The connecting subgraph tC^t is defined as the union of the shortest paths between every pair of entities in ℰ(t)E(S^t), computed by breadth-first search over the entity-level adjacency of stG_s^t. tC^t preserves the structural context that relates the rare quintuples to one another within the snapshot. Weak Signal. A weak signal is a rare subgraph t⊆ℱtS^t ^t whose content advances a tension already present in a rare subgraph t′⊆ℱt′S^t ^t at a prior observation time t′<t <t. A weak signal is therefore characterized by two properties: rarity with respect to the dominant pattern at time t, and corroboration across consecutive snapshots. Throughout this paper, the term quintuple is used as the base unit of the DTKG, where the validity period (tstart,tend)(t_start,t_end) may or may not be defined depending on whether the underlying fact carries explicit temporal bounds. The DTKG is further considered as a memory in which quintuples carry updatable properties. At each time t, the framework operates in two stages. 3.2 Module 1: Rare Subgraphs Extractor The TKG snapshot is read from the DTKG. It is then prompted to the LLM, with all quintuples presented as an indexed list of the form (subject entity name: subject entity type ⟶ predicate (tstart,tend)(t_start,t_end) ⟶ object entity name: object entity type). Following the CoT technique [24], the LLM reasons in two steps. In the first step, a baseline narrative is established: a summary of the TKG snapshot. In the second step, the quintuples are compared against the baseline narrative to identify rare subgraphs. Deviations arising solely from TKG artefacts, such as label mismatches, that are otherwise aligned with the baseline are explicitly excluded. Once the rare subgraphs have been identified, a connecting subgraph is constructed spanning them. It preserves the structural context that ties the rare subgraphs together while the remainder of the snapshot is discarded. The rare subgraphs and their connecting subgraph at time t are highlighted in memory as prior evidence for the next module. 3.3 Module 2: Weak Signal Alerter The connecting subgraph produced by Module 1 is not yet a weak signal. A single deviation observed at time t may fade away or grow into a consequential pattern in later snapshots. The purpose of this module is therefore to decide, by comparing the current connecting subgraph with those of previous snapshots, which subgraph should be labeled as weak signals. The connecting subgraphs of all prior snapshots and the current connecting subgraph tC^t are retrieved from the DTKG. If tC^t is empty, the module is skipped. Otherwise, each subgraph is prompted in the same textual form as in Module 1. Following the CoT technique [24], the LLM reasons in two steps. In the first step, for each past connecting subgraph, the LLM identifies what the snapshot was mainly about and what its rare quintuples were hinting at. In the second step, each quintuple of tC^t is compared against those past hints to decide whether its content continues the same thread. A subgraph is labeled as a weak signal if its content advances a tension already present in a prior connecting subgraph, making the earlier rare subgraph reinterpretable as a precursor of a future event. Every labeled subgraph as a weak signal is annotated and written back to memory. At the next time step, this weak signal and the related connecting subgraphs are discarded and are not shown again. Both modules are applied to every subsequent snapshot as shown in Figure 2. Figure 2: The architecture of our framework 4 Experiments Our evaluation addresses the following research questions: RQ1: How accurately does our framework detect weak signals compared to keyword-, topic-, and graph-based methods? RQ2: How early are weak signals detected relative to their establishment as strong signals, and how does lead time vary across methods? RQ3: Does the framework yield interpretable, contextually grounded explanations supporting its self-interpretable design? RQ4: Which components of the framework account for its performance? 4.1 Metrics Let =s1,…,smS=\s_1,…,s_m\ denote the set of strong signals. Each signal s∈s is characterized by two ground-truth timestamps: tsweakt^weak_s, the earliest date at which s qualifies as a weak signal, and tsstrongt^strong_s, the date at which it becomes a strong signal. The detection window of s is s=[tsweak,tsstrong]W_s=[\,t^weak_s,\,t^strong_s\,]. Each method M produces a set of detections M=(s^,t)D_M=\( s,t)\, where s s is the textual or structural description of the detected weak signal and t is its detection timestamp. Each signal s is further associated with a set (s)A(s) of discriminative anchor words derived from its ground-truth precursor facts. Detection accuracy (RQ1 and RQ4). Because the output representations of the baseline methods differ in granularity: single keywords, bags of topic words, graphlets from which the entities of the underlying graph are collected, and typed quintuples together with a natural-language interpretation, we parameterize matching by a strictness level k∈1,2,3k∈\1,2,3\. A detection (s^,t)∈M( s,t) _M is a true positive at level k if (i) the description s s shares at least k distinct words with (s)A(s) for some s∈s , and (i) t∈st _s. Let M+(k)⊆S_M^+(k) be the set of strong signals for which M produces at least one true positive at level k. Precision, recall, and F1F_1 are reported for each k: PM(k)=|TPM(k)||M|,RM(k)=|M+(k)|||,P_M(k)= |TP_M(k)||D_M|, R_M(k)= |S_M^+(k)||S|, (1) F1,M(k)=2⋅PM(k)⋅RM(k)PM(k)+RM(k).F_1,M(k)= 2· P_M(k)· R_M(k)P_M(k)+R_M(k). (2) Note that |M||D_M| is constant across k: stricter matching reduces true positives but does not change how many detections a method emits. Lead time (RQ2). For each signal s∈M+(k)s _M^+(k), let tM,sdet(k)t^det_M,s(k) be the earliest detection timestamp produced by M for s at level k. The lead time of M on s is: LeadTimeM(s,k)=tsstrong−tM,sdet(k).LeadTime_M(s,k)=t^strong_s-t^det_M,s(k). (3) We report the mean lead time L¯M(k)=1|M+(k)|∑s∈M+(k)LeadTimeM(s,k) L_M(k)= 1|S_M^+(k)| _s _M^+(k)LeadTime_M(s,k) together with the full per-signal distribution. Lead time captures the early-warning value that PMP_M and RMR_M alone cannot: a detection at t=tsstrong−1t=t^strong_s-1 is technically correct but operationally useless. Interpretability (RQ3). We provide a qualitative comparison of the outputs produced by each method on the same event. 4.2 Datasets and baseline methods No annotated benchmark currently exists for weak signal detection [6]. State-of-the-art methods are typically evaluated qualitatively by domain experts, and no quantitative gold standard is available for comparison. We introduce what is, to the best of our knowledge, the first benchmark for weak signal detection. Constructing such a benchmark presents two main challenges. First, weak signal detection requires access to the complete temporal evolution of a signal. In practice, however, only the resulting strong signal is typically documented, while the weak signals and corroborating events that precede it are rarely recorded. Second, the benchmark must be built from a domain that captures the dynamic and temporal nature of real-world phenomena. We therefore select the corporate domain, as company histories provide heterogeneous, time-stamped event records spanning multiple years. The dataset, referred to as Wiki-OpenAI, is constructed from all monthly edits of the Wikipedia page of OpenAI, collected over the period from January 2015 to December 2025. Atomic facts are extracted from each edit using ATOM [14]. Because the same fact may be reproduced across several edits, duplicates are removed. Atomic facts are grouped by the year of the source event, so the observation timestamps of the DTKG are the 11 years 2015-2025. After deduplication, 757 atomic facts are retained from the Wikipedia corpus. The Wikipedia record does not capture the complete pre-signal history of the strong signals annotated in this benchmark: some facts were present in the real world but absent from the encyclopedic record at the time of their occurrence, and are missing from the snapshots. To address this limitation, 16 atomic facts are added manually. These facts are sourced from established news media (e.g., The New York Times), and the complete list is provided in the GitHub repository. In total, the benchmark contains 773 atomic facts (757 Wikipedia-sourced and 16 manually added). Five strong signals are annotated. Each atomic fact is assigned one of four roles: background (no precursor relationship), weak_signal (a pre-signal fact with a documented lead time to a strong signal), corroboration (a pre-signal fact that reinforces a prior weak signal), or strong_signal (the anchor event). Post-signal facts are excluded from the scoreable set and are not included in the benchmark. The five strong signals, together with the date of the earliest precursor identified and the date of the strong signal itself, are listed in Table 1. An overview of the full dataset is provided in Table 2. The DTKG is constructed from these atomic facts using ATOM [14], with the hyperparameters set to their default values. Three baselines are evaluated: Yoon [26], BERTrend [6] and BEAM [1]. gpt-5.4-mini-2026-03-1722 2 https://developers.openai.com/api/docs/models/gpt-5.4-mini is chosen thanks to its low cost and performance. Because the benchmark is drawn from a widely documented organization, an LLM may have encountered its history during pretraining. We therefore derive a second variant, Wiki-OpenAI-Anon, in which every named entity is replaced by an anonymized placeholder (OpenAI → OrgVrein, Pentagon → GovLaethas), applying the same substitution to the anchor-word sets (s)A(s). Graph structure, relation types, and timestamps are unchanged. Table 1: Strong signals annotated in Wiki-OpenAI. The S date corresponds to testt_est, the date on which the event was publicly confirmed. ID Description tsweakt^weak_s tsstrongt^strong_s S_FORPROFIT_2019 Conversion to capped-profit (OpenAI LP) Dec 2015 Mar 2019 S_BOARD_COUP_2023 Altman removal and reinstatement Mar 2019 Nov 2023 S_NYT_LAWSUIT_2023 NYT v. OpenAI copyright suit Jun 2021 Dec 2023 S_DEFENCE_TURN_2025 Pivot to defence and national security Jul 2019 Jun 2025 S_FORPROFIT_PBC_2025 Conversion to for-profit PBC Dec 2017 Oct 2025 Table 2: Summary statistics of the Wiki-OpenAI benchmark dataset. Property Value Corpus Source Wikipedia (OpenAI page) Period Jan 2015 – Dec 2025 Number of yearly snapshots 11 Atomic facts Total 773 Wikipedia-sourced 757 Manually added 16 Annotation (by role) Background 736 Weak signal 22 Corroboration (pre-signal) 10 Strong signal 5 4.3 Results RQ1: Detection Accuracy As reported in Table 3, C-Unseen obtains the highest F1,M(k)F_1,M(k) score at k=2k=2 and k=3k=3, and covers all five strong signals at k=1k=1. At k=1k=1, BEAM obtains higher precision and F1F_1. All three baselines degrade as k increases: BEAM covers no signal at k=3k=3, leaving F1F_1 undefined, while BERTrend retains 1.40±0.551.40± 0.55 covered signals. Yoon is structurally unable to satisfy higher thresholds, since its detections consist of a single term. The precision and recall of C-Unseen also decrease with k, indicating that some detections do not contain sufficient anchor words and therefore constitute false positives; its coverage nevertheless remains at 4.20±1.104.20± 1.10 signals at k=3k=3, against 1.40±0.551.40± 0.55 for the best baseline. Table 4 reports the same evaluation on Wiki-OpenAI-Anon. The ranking is unchanged at every threshold, and the degradation of C-Unseen is limited: F1F_1 decreases from 0.6130.613 to 0.6030.603 at k=1k=1 and from 0.4320.432 to 0.4000.400 at k=2k=2, with coverage unchanged at five and 4.80±0.454.80± 0.45 signals respectively. C-Unseen does not detect a weak signal by recognising which company or person it concerns, but by finding facts that conflict with the rest of the snapshot and reappear in later ones. Replacing the entity names leaves both of these intact, which is why performance barely changes. The small residual difference is expected: anonymization removes not only any prior knowledge of the organization, but also the general knowledge that makes a conflict legible. For example, the defence signal rests on the tension between a stated military-use restriction and the interest shown by the Pentagon; once the entity is rendered as GovLaethas, nothing indicates that it is a defence agency, and the conflict is hard to see from the snapshot alone. These results confirm that the proposed framework detects weak signals more accurately than all evaluated baselines, which addresses RQ1. Table 3: Detection accuracy on Wiki-OpenAI at shared-anchor-word match thresholds k∈1,2,3k∈\1,2,3\. C-Unseen and BERTrend are reported as mean ± sample std over n=5n=5 runs; Yoon and BEAM are deterministic and reported from a single run. “–” denotes an undefined value. k Method |M||D_M| |M+(k)||S_M^+(k)| PM(k)P_M(k) RM(k)R_M(k) F1,M(k)F_1,M(k) 1 BEAM 19 3 0.789 0.600 0.682 BERTrend 44.8 ± 5.2 4.20 ± 0.45 0.161 ± 0.021 0.840 ± 0.089 0.270 ± 0.034 Yoon 311 3 0.032 0.600 0.061 C-Unseen 50.2 ± 10.1 5.00 0.444 ± 0.066 1.000 0.613 ± 0.061 2 BEAM 19 1 0.053 0.200 0.083 BERTrend 44.8 ± 5.2 2.80 ± 1.30 0.081 ± 0.026 0.560 ± 0.261 0.141 ± 0.047 Yoon 311 – – – – C-Unseen 50.2 ± 10.1 4.80 ± 0.45 0.281 ± 0.048 0.960 ± 0.089 0.432 ± 0.059 3 BEAM 19 0 0.000 0.000 – BERTrend 44.8 ± 5.2 1.40 ± 0.55 0.043 ± 0.030 0.280 ± 0.110 0.074 ± 0.050 Yoon 311 – – – – C-Unseen 50.2 ± 10.1 4.20 ± 1.10 0.162 ± 0.053 0.840 ± 0.219 0.271 ± 0.084 Table 4: Detection accuracy on Wiki-OpenAI-Anon. Conventions as in Table 3. k Method |M||D_M| |M+(k)||S_M^+(k)| PM(k)P_M(k) RM(k)R_M(k) F1,M(k)F_1,M(k) 1 BEAM 20 3 0.750 0.600 0.667 BERTrend 49.8 ± 2.0 4.00 0.120 ± 0.035 0.800 0.207 ± 0.053 Yoon 273 3 0.022 0.600 0.042 C-Unseen 58.4 ± 19.9 5.00 0.434 ± 0.076 1.000 0.603 ± 0.073 2 BEAM 20 1 0.050 0.200 0.080 BERTrend 49.8 ± 2.0 3.20 ± 0.45 0.092 ± 0.028 0.640 ± 0.089 0.159 ± 0.044 Yoon 273 – – – – C-Unseen 58.4 ± 19.9 4.80 ± 0.45 0.257 ± 0.066 0.960 ± 0.089 0.400 ± 0.072 3 BEAM 20 0 0.000 0.000 – BERTrend 49.8 ± 2.0 1.80 ± 0.45 0.036 ± 0.009 0.360 ± 0.089 0.066 ± 0.016 Yoon 273 – – – – C-Unseen 58.4 ± 19.9 4.20 ± 1.10 0.115 ± 0.057 0.840 ± 0.219 0.196 ± 0.086 RQ2: Lead Time As reported in Table 5, C-Unseen achieves a mean lead time of L¯Myears(k=3)=1.00 L^years_M(k=3)=1.00 year, compared to 1.70±0.451.70± 0.45 years for BERTrend. BEAM produces no covered signal at k=3k=3, rendering its mean lead time undefined. Yoon is not adapted at k=3k=3. This result should, however, be interpreted alongside the recall results in Table 3: the lead time of BERTrend is computed over 1.40±0.551.40± 0.55 covered signals, whereas the lead time of C-Unseen is computed over 4.20±1.104.20± 1.10. A mean lead time derived from one or two detections may reflect particularly early-detectable signals rather than a consistent early-warning capacity, and a method that covers the remaining signals necessarily includes those detected later. The mean lead time of C-Unseen, combined with its broader signal coverage, therefore indicates that the framework detects weak signals consistently across signals of varying detectability. This result addresses RQ2. Table 5: Mean lead time L¯Myears(k=3) L_M^years(k=3) over covered signals. “–” denotes no covered signal (mean undefined). Method Wiki-OpenAI Wiki-OpenAI-Anon BEAM – – BERTrend 1.70 ± 0.45 2.00 Yoon – – C-Unseen 1.00 1.13 ± 0.30 RQ3: Interpretability Figure 3 provides a qualitative comparison of method outputs for the signal S_DEFENCE_TURN_2025, which Yoon, BERTrend, and C-Unseen detect but BEAM does not. Yoon identifies a single keyword, which does not provide context for why it constitutes a weak signal related to the defence pivot. BERTrend produces a bag of words that captures a relevant thematic area but does not articulate the relational structure behind the signal nor the reasoning that connects it to the strong signal. In contrast, C-Unseen produces the weak signal as a subgraph of a TKG snapshot, exposing the structural tensions. This subgraph is accompanied by a natural language interpretation that traces the signal’s evolution across snapshots, identifying the transition from an abstract safety boundary to an overt tension between policy and real-world defence interests. These results confirm that the proposed framework produces self-interpretable outputs, which answers RQ3. Figure 3: Qualitative comparison of the baseline methods’ output over the detected S_DEFENCE_TURN_2025 by Yoon, BERTrend, and C-Unseen. It was not detected by BEAM. RQ4: Ablation Table 6 compares four arms against the full framework at k=3k=3. Local arms see only the quintuples of the current snapshot; full DTKG arms additionally see those of all prior snapshots. Flat arms receive a flat prompt (e.g., detect weak signals in the DTKG quintuples below), while CoT arms receive the two-step narrative-then-deviation prompt. ΔF1 F_1 is the per-run difference against C-Unseen. CoT reasoning improves on a flat prompt for the local snapshot arms, whereas extending the input to the full DTKG does not add a measurable gain over the corresponding local snapshot arm. Selecting the rare subgraphs and their connecting subgraph instead of prompting the complete graph yields ΔF1=+0.141±0.093 F_1=+0.141± 0.093 over the full DTKG arm. This addresses RQ4: the gain comes from the combination of the selected rare subgraphs and the CoT prompt, not from the LLM alone. Table 6: Ablation on Wiki-OpenAI at k=3k=3, mean ± sample std over n=5n=5 runs. Arm |M||D_M| |M+||S_M^+| PMP_M F1,MF_1,M ΔF1 F_1 local+flat 144.8 ± 18.5 5.00 0.059 ± 0.013 0.112 ± 0.023 −-0.159 local+CoT 135.2 ± 19.9 4.80 ± 0.45 0.072 ± 0.010 0.134 ± 0.018 −-0.137 ± 0.089 full DTKG+flat 195.0 ± 14.1 5.00 0.061 ± 0.009 0.115 ± 0.016 −-0.156 full DTKG+CoT 189.6 ± 10.5 5.00 0.069 ± 0.013 0.130 ± 0.023 −-0.141 ± 0.093 C-Unseen 50.2 ± 10.1 4.20 ± 1.10 0.162 ± 0.053 0.271 ± 0.084 — 5 Conclusion In this paper, we introduced C-Unseen, a self-interpretable framework for weak signal detection in Dynamic Temporal Knowledge Graphs, along with the formal definition of weak signals in this setting. The DTKG served three complementary roles in the framework: as a storage of domain knowledge over time, as a persistent memory in which the intermediate reasoning steps of the LLM are stored across snapshots, and as a structural backbone that expands the context available to the Weak Signal Alerter by connecting rare subgraphs. Experimental evaluation on the Wiki-OpenAI benchmark demonstrated that the framework achieves better performance, with self-interpretable outputs, and that this performance is retained on a named entity anonymized variant of the benchmark. Several directions are identified for future work. First, the scalability of the Rare Subgraphs Extractor to large DTKGs remains to be addressed: as the number of quintuples per snapshot grows, the full snapshot may exceed the LLM’s context window. A direction is to prompt the model over a compressed representation of the TKG snapshot, guided toward subgraph regions that are likely to contain rare subgraphs, rather than presenting the complete snapshot. Second, the DTKG in this work is constructed in a few-shot manner using ATOM; the introduction of a domain ontology to guide ATOM during construction is expected to reduce false positives. Third, the benchmark introduced here is limited to a single organization; evaluation across multiple domains would enhance the generalizability of the framework. References [1] H. Abou Jamra, M. Savonnet, and E. Leclercq (2022) BEAM: a network topology framework to detect weak signals. International Journal of Advanced Computer Science and Applications 13 (4). Cited by: §2, §4.2. [2] H. Abou Jamra, M. Savonnet, and E. Leclercq (2022) Identification of weak signals in a temporal graph of social interactions. In Proceedings of the 26th International Database Engineered Applications Symposium, IDEAS ’22, New York, NY, USA, p. 34–42. External Links: ISBN 9781450397094, Document Cited by: §1. [3] T. Ahlqvist and T. Uotila (2020) Contextualising weak signals: towards a relational theory of futures knowledge. Futures 119, p. 102543. External Links: Document Cited by: §1. [4] H. I. Ansoff (1975) Managing strategic surprise by response to weak signals. California management review 18 (2), p. 21–33. Cited by: §1. [5] C. Bisson and Ö. Y. Diner (2017) Strategic early warning system for the french milk market: a graph theoretical approach to foresee volatility. Futures 87, p. 10–23. Cited by: §2. [6] A. Boutaleb, J. Picault, and G. Grosjean (2024) BERTrend: neural topic modeling for emerging trends detection. In Proceedings of the Workshop on the Future of Event Detection (FuturED), J. Tetreault, T. H. Nguyen, H. Lamba, and A. Hughes (Eds.), Miami, Florida, USA, p. 1–17. External Links: Document Cited by: §1, §2, §4.2, §4.2. [7] B. Cai, Y. Xiang, L. Gao, H. Zhang, Y. Li, and J. Li (2023) Temporal knowledge graph completion: a survey. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI ’23. External Links: ISBN 978-1-956792-03-4, Document Cited by: §1. [8] Y. Cao, L. Wang, and L. Huang (2025) DPCL-diff: temporal knowledge graph reasoning based on graph node diffusion model with dual-domain periodic contrastive learning. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence, AAAI’25/IAAI’25/EAAI’25. External Links: ISBN 978-1-57735-897-8, Document Cited by: §2. [9] F. Dotsika and A. Watkins (2017) Identifying potentially disruptive trends by means of keyword network analysis. Technological forecasting and social change 119, p. 114–127. Cited by: §2. [10] A. Ebadi, A. Auger, and Y. Gauthier (2026) WISDOM: an ai-powered framework for emerging research detection using weak signal analysis and advanced topic modelling. Journal of Informetrics 20 (1), p. 101759. External Links: ISSN 1751-1577, Document Cited by: §1. [11] M. El Akrouchi, H. Benbrahim, and I. Kassou (2021) End-to-end lda-based automatic weak signal detection in web news. Knowledge-Based Systems 212, p. 106650. External Links: ISSN 0950-7051, Document Cited by: §2. [12] E. Hiltunen (2008) The future sign and its three dimensions. Futures 40 (3), p. 247–260. External Links: Document Cited by: §1. [13] L. Kwon, J. Park, Y. Moon, B. Lee, Y. Shin, and Y. Kim (2018) Weak signal detecting of industry convergence using information of products and services of global listed companies-focusing on growth engine industry in south korea–. Journal of Open Innovation: Technology, Market, and Complexity 4 (1), p. 1–19. Cited by: §2. [14] Y. Lairgi, L. Moncla, K. Benabdeslem, R. Cazabet, and P. Cléau (2026) ATOM: AdapTive and OptiMized dynamic temporal knowledge graph construction using LLMs. In Findings of the Association for Computational Linguistics: EACL 2026, V. Demberg, K. Inui, and L. Marquez (Eds.), Rabat, Morocco, p. 950–966. External Links: Document, ISBN 979-8-89176-386-9 Cited by: §1, §1, §3.1, §4.2, §4.2. [15] Y. Lairgi, L. Moncla, R. Cazabet, K. Benabdeslem, and P. Cléau (2024) Itext2kg: incremental knowledge graphs construction using large language models. In International Conference on Web Information Systems Engineering, p. 214–229. Cited by: §1. [16] Y. Lee and J. Park (2018) Identification of future signal based on the quantitative and qualitative text mining: a case study on ethical issues in artificial intelligence. Quality & Quantity 52 (2), p. 653–667. Cited by: §2. [17] Z. Li, X. Jin, W. Li, S. Guan, J. Guo, H. Shen, Y. Wang, and X. Cheng (2021) Temporal knowledge graph reasoning based on evolutional representation learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, p. 408–417. Cited by: §2. [18] R. Liao, X. Jia, Y. Li, Y. Ma, and V. Tresp (2024) GenTKG: generative forecasting on temporal knowledge graph with large language models. In Findings of the Association for Computational Linguistics: NAACL 2024, K. Duh, H. Gomez, and S. Bethard (Eds.), Mexico City, Mexico, p. 4303–4317. External Links: Document Cited by: §2. [19] J. Maitre, M. Menard, G. Chiron, and A. Bouju (2019) Détection de signaux faibles dans des masses de données faiblement structurées. Recherche d’Information, Document et Web Sémantique 3 (1). Cited by: §2. [20] S. Niu, K. Yang, R. Zhao, Y. Liu, Z. Li, H. Wang, and W. Chen (2025) Tree-KG: an expandable knowledge graph construction framework for knowledge-intensive domains. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar (Eds.), Vienna, Austria, p. 18516–18529. External Links: Document, ISBN 979-8-89176-251-0 Cited by: §1. [21] C. Park and S. Cho (2017) Future sign detection in smart grids through text mining. Energy Procedia 128, p. 79–85. Cited by: §2. [22] L. Pépin, P. Kuntz, J. Blanchard, F. Guillet, and P. Suignard (2017) Visual analytics for exploring topic long-term evolution and detecting weak signals in company-targeted tweets. Computers & Industrial Engineering 112, p. 450–458. Cited by: §2. [23] J. Wang, K. Sun, L. Luo, W. Wei, Y. Hu, A. W. Liew, S. Pan, and B. Yin (2024) Large language models-guided dynamic adaptation for temporal knowledge graph reasoning. In Advances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (Eds.), Vol. 37, p. 8384–8410. External Links: Document Cited by: §2. [24] J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V. Le, D. Zhou, et al. (2022) Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, p. 24824–24837. Cited by: §3.2, §3.3. [25] Y. Xia, D. Wang, Q. Liu, L. Wang, S. Wu, and X. Zhang (2024) Chain-of-history reasoning for temporal knowledge graph forecasting. In Findings of the Association for Computational Linguistics: ACL 2024, L. Ku, A. Martins, and V. Srikumar (Eds.), Bangkok, Thailand, p. 16144–16159. External Links: Document Cited by: §1. [26] J. Yoon (2012) Detecting weak signals for long-term business opportunities using text mining of web news. Expert Systems with Applications 39 (16), p. 12543–12550. External Links: Document Cited by: §1, §2, §4.2. [27] L. Zhong, J. Wu, Q. Li, H. Peng, and X. Wu (2023) A comprehensive survey on automatic knowledge graph construction. ACM Comput. Surv. 56 (4). External Links: ISSN 0360-0300, Document Cited by: §1.