This study presents an innovative framework for predicting school success, leveraging Large Language Models (LLMs) to define ground truth labels based on comprehensive school information with Factor-Reasoning-Classification (FRC) prompting. In this article, we conceptualize school data as a complex social network and create two different graphs where schools are defined as nodes. Similarity Graph captures school similarities, integrating factors such as graduation rates, ACT scores, socioeconomic conditions, crime rates, and community resources. Geographic Proximity Graph models spatial relationships among schools using geographical coordinates. We define Merged GNN that enhances prediction accuracy by incorporating both similarity-based and spatial proximity-based information. Our approach leverages Graph Neural Networks (GNN) to predict the most probable labels the LLM model identifies. Experimental results on the school success dataset not only demonstrate the superior predictive performance of our methodology over baseline models but also highlight the importance of integrating diverse sources of information for accurate prediction and analysis.