Weak signals refer to early, low‑visibility indicators that appear before a major change becomes established. Existing methods rely on keyword frequency, topic modeling, or untyped graph topology, which struggle to capture the semantic and relational patterns underlying such signals. This paper introduces C‑Unseen, a self‑interpretable framework designed for weak signal detection in Dynamic Temporal Knowledge Graphs (DTKGs). We define a weak signal as a rare, semantically coherent subgraph that propagates across consecutive TKG snapshots. The framework consists of two modules:
Rare Subgraph Extractor: A large language model (LLM) performs chain‑of‑thought reasoning on each snapshot to identify subgraphs whose content conflicts with the dominant narrative. The LLM is prompted with “List subgraphs that are inconsistent with the main storyline of the current snapshot and explain why,” producing a candidate set $\mathcal{R}_t$.
Weak Signal Alerter: The persistence of $\mathcal{R}_t$ over time is tracked. If a subgraph appears in at least $k$ consecutive snapshots, it is flagged as a weak signal. Formally: $$ \text{WeakSignal}(S)= \begin{cases} 1, & \text{if } \forall t\in[T,T+k-1],\; S\in\mathcal{R}_t \ 0, & \text{otherwise} \end{cases} $$
Experiments on several public DTKG datasets compare C‑Unseen against keyword‑, topic‑, and graph‑based baselines. C‑Unseen achieves notable gains in both recall and precision, especially in noisy environments.
Blogger's Review: The framework cleverly leverages LLM reasoning together with temporal graph dynamics, offering an interpretable approach to weak signal detection that merits further testing in real‑world intelligence systems.