Business process remaining time prediction using explainable reachability graph from gated RNNs.

Autor: Cao, Rui, Zeng, Qingtian, Ni, Weijian, Duan, Hua, Liu, Cong, Lu, Faming, Zhao, Ziqi
Předmět:
Zdroj: Applied Intelligence; Jun2023, Vol. 53 Issue 11, p13178-13191, 14p
Abstrakt: Gated recurrent neural networks (RNNs) are successfully applied to predict the remaining time of business processes. Existing methods typically train multiple prediction models for prefixes bucketing. Furthermore, the gated RNNs are more like black boxes and lack interpretability. An explainable gated RNN using a reachability graph is proposed to improve the results of prediction. First, prefixes of the event log are generated to train a single prediction model, and hidden states of the gated RNN are saved. Second, a Petri net and its corresponding reachability graph are constructed by taking an event log as input. Next, the hidden states of the gated RNN are mapped to a reachability state of the reachability graph by the decoding mapping to explain the remaining time prediction model, i.e., gated RNN. Finally, our method is validated by six real-life event logs. Based on the experimental results, it is shown that a mapping from the hidden states of the gated RNN to a reachability state of the reachability graph is established, and the gated RNN that recognizes transition sequences is also explained for improving the performance of remaining time prediction of business processes. [ABSTRACT FROM AUTHOR]
Databáze: Complementary Index
Externí odkaz: