Sequential Attention Source Identification Based on Feature Representation

TL;DR

This work tackles rumor source localization under time-varying infection with heterogeneous user interactions. It introduces TGASI, a transferable sequence-to-sequence framework that combines a GNN encoder (learning a heterogeneous influence matrix and dynamic/topological features) with a Bi-GRU decoder augmented by a temporal attention mechanism to identify sources across timestamps. The authors design a graph-aware loss and conduct comprehensive experiments on six real networks, showing that TGASI consistently outperforms state-of-the-art methods and demonstrates inductive transferability to unseen networks and diffusion models. The approach advances practical source localization by enabling accurate, scalable predictions in diverse propagation scenarios with limited timestamped observations.

Abstract

Snapshot observation based source localization has been widely studied due to its accessibility and low cost. However, the interaction of users in existing methods does not be addressed in time-varying infection scenarios. So these methods have a decreased accuracy in heterogeneous interaction scenarios. To solve this critical issue, this paper proposes a sequence-to-sequence based localization framework called Temporal-sequence based Graph Attention Source Identification (TGASI) based on an inductive learning idea. More specifically, the encoder focuses on generating multiple features by estimating the influence probability between two users, and the decoder distinguishes the importance of prediction sources in different timestamps by a designed temporal attention mechanism. It's worth mentioning that the inductive learning idea ensures that TGASI can detect the sources in new scenarios without knowing other prior knowledge, which proves the scalability of TGASI. Comprehensive experiments with the SOTA methods demonstrate the higher detection performance and scalability in different scenarios of TGASI.

Figures (4)

• Figure 1: The illustration of the multiple rumor source detection problem under available snapshots with discrete timestamps. Due to the constraints of time and space consumption, the timestamps of snapshots we can capture are limited and discrete.
• Figure 2: The illustration of TGASI based on the sequence-to-sequence framework. The features generation encoder includes four innovative modules. (a) influence probability transition matrix $\bm{\mathcal{W}}$ is estimated by available captured snapshots. (b) coarse-grained source probability feature in one timestamp $s$ is designed based on the influence matrix $\bm{\mathcal{W}}$. (c) dynamic infected feature $H_2$ and uninfected feature $H_3$ is generated from each node's neighbors. (d) lower dimensional embedding $H_G$ of the topology structure is obtained based on the topology graph. The GRU-based decoder with temporal attention uses Bi-GRU to decode the embedding, then (e) a one-timestamp based attention mechanism is designed for time series information in order to distinguish and weight the importance of decoder information at each different timestamp for source localization task. What's more, (f) a graph constraint based loss function is specially designed for the localization task to train TGASI.
• Figure 3: The performance evaluation of variant models from TGASI in $G_2$ to $G_4$.
• Figure 4: The performance evaluation of TGASI on the inductive learning task. (a)-(c) is the inductive learning on different social networks based on the IC model, and (d)-(f) is on different propagation models in $G_3$. $P1$, $P2$, $P3$, and $P4$ are denoted as homogeneous SI, homogeneous SIR, heterogeneous SI, and heterogeneous SIR.

Theorems & Definitions (1)

• Theorem 1