Insider threats represent a significant security challenge within organizations due to their often stealthy nature and the inherent trust placed in internal personnel. To address this complex issue, this study explores the efficacy of advanced graph-based neural network models in identifying potential insider threats from organizational data. Using the CMU CERT 4.2 dataset, we first constructed a comprehensive knowledge graph from existing datasets to model the intricate relationships and interactions among employees. Subsequently, we employ three distinct graph convolutional network (GCN) architectures: a standard GCN, a Spatio-Temporal GCN (STGCN), and a Capsule GNN (Graph Neural Network). Our experiments are methodically conducted across varying sizes of dataset, focusing on detection accuracy as a primary metric. The performance of each model is rigorously compared, with the STGCN consistently demonstrating superior accuracy of roughly 94% while maintaining lower false positive rates across all datasets tested. The findings indicate that incorporating both spatial and temporal data dimensions via STGCN significantly enhances the ability to detect and predict insider threats more effectively than traditional GCN and Capsule GNN architectures. This research underscores the potential of spatio-temporal graph analytics as a promising approach in the proactive management of internal security risks.