As the influence of social networks on consumer behavior becomes increasingly significant, existing studies often ignore the dynamics of network topology and the heterogeneity of user interactions when modeling the propagation process, resulting in insufficient prediction accuracy of the behavior diffusion path and influence. To this end, this paper proposes a consumer behavior propagation model based on graph neural network (GNN-CPD), which aims to reveal the propagation mechanism of consumer decisions in social networks through dynamic graph structure learning and high-order relationship reasoning. This method first uses Graph Attention Network (GAT) and GraphSAGE to construct a two-layer message passing framework to capture the heterogeneous characteristics of user nodes and the dynamic influence weights of neighbor nodes, respectively. Secondly, it introduces the Temporal Convolutional Module (TCN) to process the spatiotemporal dependencies of user behavior sequences, and designs a joint loss function to simultaneously optimize the two tasks of consumer behavior classification and propagation path prediction. Experiments on Weibo e-commerce datasets show that GNN-CPD significantly outperforms existing models in terms of consumer behavior prediction accuracy and key user identification F1-score, especially in high-density social subgraphs (node degree>50), where the lowest conversion rate prediction error is 2.83%. This model can effectively quantify the nonlinear impact of social relationship strength on consumer decisions, and provides an interpretable topological feature analysis tool for seed user selection in precision marketing.