Task-Free Continual Learning with Dynamic Loss for Online Next Activity Prediction
Tamara Verbeek, Ruozhu Yao, Marwan Hassani
Abstract
Abstract Continual learning, known also as lifelong learning, aims at designing learning models that can continuously and autonomously adapt to varying data concepts without forgetting previously collected knowledge. Such concepts are referred to as tasks. Predictive business process monitoring, which predicts future process steps, is crucial in dynamic environments where tasks are not previously specified and processes frequently change or face unpredictability. However, many existing frameworks assume a static setting, ignoring dynamic nature and concept drifts in processes, leading to catastrophic forgetting—where training over new data adversely affects the performance on previously learned tasks. This paper presents TFCLPM, a framework for online next activity prediction that operates without relying on predefined tasks and employs continual learning techniques to reduce catastrophic forgetting. The methodology combines a Single Dense Layer neural network with a continual learning algorithm designed to retain challenging historical samples and include a regularizer to stabilize model parameters. Extensive experimental evaluations with synthetic and real-world event logs highlight our optimal configurations. The proposed framework’s performance is compared against three existing online next activity prediction methodologies. Results show significant improvements in prediction accuracy, especially in scenarios with gradual or recurrent drifts, highlighting the framework’s robustness and efficiency, even with large datasets.