Predicting User Risk Behavior in the Metaverse Using Hybrid Bidirectional LSTM–GRU Temporal Deep Learning

Abstract

The rapid expansion of metaverse environments has created a complex ecosystem where user interactions generate vast amounts of behavioral and transactional data. Understanding and predicting user risk behavior has become a critical challenge for maintaining trust, safety, and economic stability within these digital spaces. This study proposes a hybrid Bidirectional Long Short-Term Memory and Gated Recurrent Unit (LSTM–GRU) deep learning model to forecast user risk behavior based on temporal patterns derived from metaverse transaction data. The model was trained on sequential features, including login frequency, session duration, and transaction amount, to capture both short-term fluctuations and long-term behavioral dependencies. Experimental evaluation demonstrated that the proposed model achieved a Root Mean Square Error (RMSE) of 7.342, a Mean Absolute Error (MAE) of 5.813, and a coefficient of determination (R²) of 0.352. The results indicate that the hybrid architecture effectively learns temporal dynamics and accurately forecasts user risk trends, providing a reliable foundation for proactive risk detection. The consistent convergence between training and validation loss further confirms the model’s stability and generalization capability. Overall, this research contributes to the integration of artificial intelligence in metaverse governance by offering a data-driven approach to behavioral forecasting that can enhance user safety, improve decision-making, and support the development of intelligent, adaptive risk management systems in virtual environments.