FGSSI: A Feature-Enhanced Framework With Transferability for Sequential Source Identification
Dongpeng Hou, Chao Gao, Zhen Wang, Xuelong Li
Abstract
Source localization based on snapshot observations has gained popularity owing to its low cost and accessibility. However, current methods do not adequately address the impact of user interaction in time-varying infection scenarios, leading to reduced accuracy in heterogeneous interaction environments. To address these issues and ensure transferability and applicability in different scenarios, this paper proposes an inductive localization framework based on sequence-to-sequence architecture, named Feature-representation-based Graph Sequential Source Identification (FGSSI). Particularly, FGSSI employs a user-feature-construction-based encoder to construct low-dimensional embeddings of each individual by evaluating their dynamic and static behavioral attributes at different timestamps. Additionally, FGSSI uses a graph-GRU based decoder with temporal attention to infer the source by incorporating graph topology information into the sequential model and handling imbalanced infection information under temporal characteristics. It is worth mentioning that FGSSI can detect sources in new scenarios without prior knowledge owing to the proposed inductive learning approach. Comprehensive experiments using SOTA methods demonstrate FGSSI's superior detection performance and transferability.