Straight-Through and Autoregressive Influence Estimation for Influence Maximization

Influence maximization(IM) aims to select a set of seed nodes in a social network to maximize the expected spread of information. It has long been a prominent research topic in statistical physics of social network analysis; however, traditional methods often rely on simplified or model-free propagation assumptions, which limit their effectiveness in real-world scenarios. Recently, learning-based approaches have emerged to better capture the complexity of real-world propagation dynamics. Despite this progress, neural networks—being optimized through gradient-based methods over continuous variables—still struggle with the combinatorial nature of the IM problem. To address this, we propose STARIM, a novel model-based optimization framework. STARIM leverages Graph Neural Network(GNN) and an autoregressive approach to model the continuous evolution of node states, and incorporates Straight-Through Estimation(STE) to enable efficient end-to-end optimization over a continuous seed space. This design allows for the inclusion of additional constraints, such as varying costs for each seed node, or targeted IM, where influence is maximized over a predefined target set. Extensive experiments on real-world datasets demonstrate that our method consistently outperforms state-of-the-art baselines, highlighting its effectiveness and robustness. Moreover, the flexibility of our framework makes it readily adaptable to a wide range of learning-based optimization problems in networked systems.