Towards Learning the Optimal Sampling Strategy for Suffix Prediction in Predictive Monitoring

Predictive monitoring is a subfield of process mining which focuses on forecasting the evolution of an ongoing process case. A related main challenge is activity suffix prediction, the problem of predicting the sequence of future activities until a case ends. One aspect that has been neglected is the activity selection strategy during inference and its impact on the results. This paper introduces the “Deep Reinforcement Learning Predictor” (DOGE), a system which leverages Deep Reinforcement Learning to learn the ideal sampling strategy during training of the neural model. This approach not only simplifies the design of the neural network but also enhances inference speed by avoiding explorative sampling strategies. Through an extensive evaluation against established benchmarks, DOGE shows significant improvements in both performance and adaptability across diverse event log characteristics, highlighting the efficacy of reinforcement learning in predictive monitoring.