TY - JOUR
T1 - SIGRL
T2 - Sociologically-Informed Graph Representation Learning for Social Influence Prediction
AU - Xu, Haowei
AU - Gao, Chao
AU - Li, Xianghua
AU - Wang, Zhen
N1 - Publisher Copyright:
© 2013 IEEE.
PY - 2025
Y1 - 2025
N2 - Social influence shapes various aspects of our lives, both in our daily physical interactions and in the virtual realm of the Web. Predicting social influence for individual users with high accuracy is essential. Traditional opinion dynamics models often depend on computer simulations, which require calibration using real-world data, resulting in significant manual effort and computational complexity. While recent deep-learning methods on graphs have demonstrated flexibility in modeling social interactions, they tend to overlook established scientific knowledge regarding social interaction mechanisms. To address these limitations, this paper introduces SIGRL, a Sociologically-Informed Graph Reresentation Learning for predicting social influence. SIGRL begins by randomly sampling a user's ego network as input. Drawing on theoretical models, we design a novel message-passing mechanism based on the bounded confidence model (BCM), refining the influence weights of local nodes during message propagation. Kolmogorov-Arnold Networks (KANs) are subsequently used to update node features, enhancing both the model's expressiveness. The theoretical model is further redefined as an Ordinary Differential Equation (ODE), incorporated as a constraint during neural network training. Extensive experiments on the Open Academic Graph (OAG), Twitter, Weibo, and Digg datasets demonstrate that SIGRL significantly outperforms existing baseline methods.
AB - Social influence shapes various aspects of our lives, both in our daily physical interactions and in the virtual realm of the Web. Predicting social influence for individual users with high accuracy is essential. Traditional opinion dynamics models often depend on computer simulations, which require calibration using real-world data, resulting in significant manual effort and computational complexity. While recent deep-learning methods on graphs have demonstrated flexibility in modeling social interactions, they tend to overlook established scientific knowledge regarding social interaction mechanisms. To address these limitations, this paper introduces SIGRL, a Sociologically-Informed Graph Reresentation Learning for predicting social influence. SIGRL begins by randomly sampling a user's ego network as input. Drawing on theoretical models, we design a novel message-passing mechanism based on the bounded confidence model (BCM), refining the influence weights of local nodes during message propagation. Kolmogorov-Arnold Networks (KANs) are subsequently used to update node features, enhancing both the model's expressiveness. The theoretical model is further redefined as an Ordinary Differential Equation (ODE), incorporated as a constraint during neural network training. Extensive experiments on the Open Academic Graph (OAG), Twitter, Weibo, and Digg datasets demonstrate that SIGRL significantly outperforms existing baseline methods.
KW - graph representation learning
KW - Influence locality prediction
KW - physics-informed neural networks
KW - social networks
UR - http://www.scopus.com/inward/record.url?scp=105005197059&partnerID=8YFLogxK
U2 - 10.1109/TNSE.2025.3569545
DO - 10.1109/TNSE.2025.3569545
M3 - 文章
AN - SCOPUS:105005197059
SN - 2327-4697
JO - IEEE Transactions on Network Science and Engineering
JF - IEEE Transactions on Network Science and Engineering
ER -