Churn prediction in the e-commerce field is very important for identifying customers at risk of leaving a platform, allowing businesses to proactively intervene and retain valuable clientele. By the advanced predictive models, e-commerce companies can optimize marketing strategies, enhance customer satisfaction, and ultimately improve long-term profitability. This study introduces an improved predictive framework tackling two primary obstacles in churn prediction within the e-commerce sector. By amalgamating a thorough comprehension of customer behaviors through RFM analysis and subsequent churn prediction via a graph-based model, the proposed method addresses the imbalance issue prevalent in such predictions. This method employs degree-based sampling alongside a max pooling aggregation function within the GraphSAGE model framework, specifically tailored to address imbalanced learning challenges. By effectively aggregating information from strategically selected neighbors and diverse relations, the model demonstrates superior performance compared to alternative graph neural network models such as GCN, GAT, and GraphSAGE. This innovative approach holds promise for enhancing the performance of churn prediction in e-commerce settings while mitigating the effects of class imbalance.