Topic-aware influence maximization with deep reinforcement learning and graph attention networks

Taha Halal, Bogdan Cautis, Benoît Groz, Ruize Gao

Abstract: Influence maximization is a fundamental problem in network analysis, focusing on identifying a subset of nodes in a social network to maximize the spread of influence. In this paper, we present an approach for tackling the Influence Maximization (IM) problem, integrating Deep Reinforcement Learning (DRL) techniques with attentive Graph Neural Networks (GATs). Our study builds upon a prior algorithm (S2V-DQN-IM) and progressively refines it towards IM-GNN, ultimately achieving competitive performance against state-of-the-art methods on classic IM. Through experiments on benchmark datasets, we empirically validate the effectiveness of graph attention mechanisms and positional encoding, using the graph magnetic Laplacian, to reach state-of-the-art performance in terms of influence spread. Building on this success, we extend our IM-GNN framework to incorporate topic-awareness in TIM-GNN, recognizing the inherent topical nature of real-world diffusions. By harnessing probabilistic techniques, we construct topic-aware social graphs using real cascades and assess the effectivenesss of TIM-GNN on them. Our extensive experimental results validate the utility of our topic-aware approach, demonstrating significant advances over existing topic-aware IM methods. Finally, in order to improve upon performance (latency) at query time, we develop a variant of TIM-GNN, called TIM-GNN\(^x\), by using cross-attention mechanisms. We show it maintains comparable overall spread performance as its predecessor, while achieving a 10x-20x speed-up.

External IDs:dblp:journals/datamine/HalalCGG25