Dynamic Adaptive Parametric Social Network Analysis using Reinforcement Learning: A Case Study in Topic-aware Influence Maximization

Current network analysis algorithms often rely on search methods or centrality measures but face challenges such as (1) The solution space is large, resulting in high computational complexity. (2) Algorithms may be instance-dependent, relying considerably on network structure and characteristics, which may result in varying performance across different networks. (3) Most existing centrality measures are inherently static, which fail to capture the dynamic nature inherent in network analysis problems. To address these issues, this paper introduces a dynamic adaptive parametric (DAP) approach using reinforcement learning. As a case study, the method has been applied to the topic-aware influence maximization (TIM) problem, where the objective is to identify k influential nodes that maximize influence spread under a given topic vector and diffusion model. The paper introduces two dynamic centrality measures that capture the evolving importance of nodes during topic propagation in the network. To avoid instance-dependence, an adaptive reinforcement learning technique is used to adjust the significance of each measure based on the current network structure, tailoring solutions to the specific network. The parametric approach further reduces the search space by transforming TIM into a parametric optimization task, where the goal is to determine the optimal importance of each centrality measure. The proposed algorithm is evaluated on both real-world and synthetic networks. Experimental results show that the method outperforms conventional centrality-based greedy algorithms and other existing approaches in terms of solution quality, running time, and scalability. Also, as a part of our research, we propose a topic-aware benchmark dataset by augmenting the Deezer music-based social network with labeled nodes and edges, providing a valuable resource for evaluating future research.