Researchers have introduced a novel artificial intelligence framework capable of forecasting criminal activity with significantly improved precision, achieving an 86.3 percent accuracy rate in predicting robberies across major United States metropolitan areas. Published in the International Journal of Innovative Computing and Applications, the system merges spatial analysis, temporal pattern recognition, and generative data modeling to overcome traditional limitations in predictive policing algorithms. The architecture integrates three distinct machine learning components. A graph convolutional network maps geographic relationships and identifies high-risk locations, while a transformer module analyzes sequential data to detect temporal trends. To address data scarcity and mitigate training instabilities such as vanishing gradients and output bias, the framework incorporates a generative adversarial network enhanced by a variational autoencoder. This hybrid design enables the model to process complex, multi-dimensional datasets that conventional statistical methods typically struggle to evaluate. During validation trials utilizing historical crime records from Los Angeles, Seattle, and other U.S. municipalities, the system outperformed existing benchmarks by a margin of over three percentage points. While the 86.3 percent accuracy figure specifically applies to robbery forecasting, the architecture demonstrated comparable reliability across additional offense categories. Law enforcement agencies could leverage these predictions to optimize patrol routing, deploy preventive measures in emerging hotspots, and allocate finite resources more efficiently. Despite the improved metrics, the researchers acknowledge notable operational constraints. Predictive performance degrades considerably in neighborhoods lacking robust historical records, limiting immediate deployment in data-sparse or newly developed regions. To address this vulnerability, the research team plans to implement transfer learning techniques, enabling the model to generalize knowledge from well-documented urban centers to under-served areas without requiring extensive localized training data. As municipal agencies increasingly evaluate algorithmic tools for public safety applications, this development underscores both the potential and the technical hurdles of next-generation predictive systems. Continued refinement will likely focus on enhancing geographic generalization, reducing algorithmic bias, and establishing transparent validation protocols to ensure reliable integration into community policing strategies.