<p>Deep learning architectures extract insights from extensive datasets, thereby enhancing human decision-making with more rational, agile, and precise predictions. Recent research in business process management has increasingly leveraged deep learning to support predictive process monitoring-encompassing tasks such as next-activity prediction, remaining time estimation, and workload forecasting. In this paper, we introduce a novel framework designed to forecast subsequent activities in the temporal sequences of individual process instances. Our method integrates a mixture-of-experts mechanism with the Switch-Transformer neural network architecture, and the model is trained from scratch using historical process enactment logs to estimate the probabilities of upcoming activities. Furthermore, the predicted activities are displayed in a temporal graph that preserves their sequential order for each running process instance. This visualization not only clarifies the predictive process monitoring paradigm and its underlying decision-making workflow but also streamlines risk assessment and analysis in business process monitoring operations. To validate the operational efficacy of our approach, we conducted extensive experiments on both synthetic and real-world process event log datasets commonly employed in previous studies. The results demonstrate that our method not only outperforms existing techniques in predictive accuracy but also offers superior usability and practical applicability in real-world settings.</p>