A Bayesian network approach for dynamic behavior analysis: Real-time intention recognition

Dynamic intention recognition is widely applied across diverse domains, including autonomous driving, e-commerce, and human–computer interaction, to understand and identify individuals’ evolving behavioral intentions. While observable behaviors often serve as proxies for underlying intentions, accurately establishing the relationships between dynamic behaviors and evolving intentions becomes a challenging task. Moreover, external factors introduce dynamism and noise into behavioral data, complicating the process of inferring intentions. To address these challenges, we propose a novel Bayesian network approach that comprehensively models the real-time inference of behavioral intentions from dynamic behaviors. Our model incorporates a filtering mechanism designed to process evolving, noisy, and time-stamped behavioral data, enhancing data quality and ensuring reliable intention recognition. By mapping latent states to intentions through conditional dependencies and visualizing the generative process using directed acyclic graphs, we provide a transparent representation of the model's structure and reasoning. Experimental evaluations conducted on both real and synthetic datasets demonstrate the superior performance of our model compared to existing benchmarks, particularly in handling imbalanced data and minority classes. Furthermore, we extend our analysis to multi-target intention recognition scenarios, validating the model's adaptability in inferring the intentions of multiple individuals concurrently. Our approach offers a practical tool for decision analysis, empowering managers and practitioners to understand, predict, and proactively respond to individual behavioral intentions, thereby facilitating the development of targeted strategies and personalized services.