Predictive business process monitoring aims to forecast future characteristics of an ongoing process by analyzing event data. Gaining foresight into process execution holds significant potential for enhancing operational efficiency, optimizing resource management, and delivering more effective customer services. Next activity (action) prediction is one of the fundamental problems in predictive process monitoring. Most research on this problem involved the use of deep-learning models. However, these black-box models often fall short in terms of explainability and temporal reasoning. In this paper, we explore the utility of Dynamic Bayesian Networks (DBN) as a white-box alternative to bridge the gap between performance and interpretability. A key challenge in using DBNs has been their struggle in handling unseen data combinations. In this paper, we introduce InferNet, a Dynamic Bayesian Network based approach that tackles unseen data combinations alongside providing the benefits of explainability of DBNs. We introduce a novel estimation approach called score-based marginalization to handle unseen data that leverages the network structure and evaluates parent contributions to improve the accuracy of posterior distribution estimation. We demonstrate that our approach significantly surpasses existing white-box methods and either outperforms or competes closely with more complex deep-learning-based black-box models while having the potential to provide explainable insights into process execution.