<p>Collusive review spammers on e-commerce platforms usually post-coordinated deceptive content at scale to manipulate ratings and mislead consumers. Existing methods for detecting review spammers mainly focus on reviewing behaviors, neglecting users’ long-term aspect preferences reflected in historical reviews. This study introduces a graph-based pairwise Markov random field (pMRF) framework for collusive spammer detection by integrating users’ behavioral features and long-term aspect preferences. Specifically, we use a large language model (LLM) to extract long-term aspect preferences and align them with behavioral features via a self-contrastive learning strategy. Moreover, we integrate a dual-level graph attention network (GAT) with edge potential learning to estimate collusive interaction strengths from both behavior and preferences, and explicitly calibrate the learned relation strengths into signed pMRF edge potentials for globally consistent and explainable propagation. We evaluate the proposed framework on two real-world datasets, AmazonCN and YelpZip, and results show that our method consistently outperforms strong baselines across multiple metrics.</p>