Predictive Business Process Monitoring Approach Based on Hierarchical Transformer

Predictive business process monitoring is predicting the next stage of based on sequence events that have occurred in instance, which positive for promoting rational allocation resources and improvement execution efficiency. There are drawbacks modeling instances, such as conceptual drift phenomenon long event sequences. Therefore, we propose a hierarchical Transformer-based prediction model to improve performance predictive model. We encode features using two different encoding methods obtain relationship between activities attributes. A detection algorithm proposed segment calculate correlation segments by cross-attention. Furthermore, learnable position designed capture relative information subsequences. Finally, granularity, attributes, subsequences, complete fused weights. Experiments were run seven real logs activity remaining time prediction, accuracy improved 6.32% average, mean absolute error reduces 21% average.