In modern society the pervasive use of credit cards has led to a substantial increase in fraudulent activities, resulting in considerable financial losses for both cardholders and issuing banks. Despite the efforts of academic and industry researchers to develop various algorithms for fraud detection, the multidimensional attributes of transaction data lack both a comprehensive exploration and applications. To bridge this gap this study introduces a model based on a stacked temporospatial graph attention residual network (stacked TS-GARN), specifically tailored for credit card fraud detection. The model first employs a temporal-spatial Node2Vec with attention (TSN2VA) method to generate embedding vectors that capture the structural information of transaction nodes in both temporal and spatial dimensions. Subsequently, it integrates these temporal and spatial embedding vectors with their corresponding attribute features and further aggregates the features of neighboring transaction nodes through a multihead attention mechanism to produce a comprehensive feature vector. This vector is then used as input for subsequent classification tasks. For fraud detection, a two-layer residual network logistic stacking ensemble (TLSE) is used. Finally, experiments are conducted on two publicly available credit card data sets to evaluate the model’s performance across various dimensions. The experimental results demonstrate that the stacked TS-GARN model effectively extracts latent features and achieves more accurate recognition outcomes than the other tested models.