Source Localization on Complex Financial Semantic Graphs
Source localization aims to locate the origin of failures and rumors in complex graphs, which has broad implications for financial systems. However, previous methods commonly assume that the propagation graph is known in advance. In practice, the graph on which failures or information propagate depends on relationships among assets and other factors that are difficult to observe. In this paper, we first construct a propagation proxy graph from the semantic similarity between financial posts by using a pretrained language model to encode their textual descriptions. We then design a two-stage framework for source localization based on a voting scheme. In the first stage, each activated node votes for the neighbor that may have triggered it. In the second stage, each node combines the votes received from its neighbors and determines whether it belongs to the source set. In the experiment, approximately 12,000 financial Twitter posts are encoded with a pretrained language model, and the propagation proxy graph is constructed by connecting the Top-k semantic neighbors of each node. The source-localization model is trained using susceptible–infected–recovered (SIR) trajectories and then applied to susceptible–infected (SI), independent cascade (IC), and linear threshold (LT) trajectories. This protocol simulates concept drift caused by a change in the underlying propagation mechanism. Across the three target diffusion settings, accuracy remains around 0.90, AUC exceeds 0.91, and macro-averaged precision ranges from 0.565 to 0.580. The experiments demonstrate that the proposed method remains effective when the diffusion mechanism changes after training.