Research On Business Process Instance Remaining Time Prediction Using Deep Learning

As a predictive business process monitoring task,remaining time prediction has become a hot research topic in the field of business process management in recent years.The aim is to predict the remaining time of the executing process instances.In existing deep learning approaches,prediction in the business process remaining time mostly utilises natural language process method,which applies the traditional prediction model of recurrent neural networks to constructing a single predictive mode.This model neglects distinguishing between the event of the business process and the phrases of the natural language sequence,thus lacks the pertinency of different types of real cases in business practice.This thesis focuses on the actual event log as the research object and design of the deep learning model combined with the characteristics of the business process.The framework of this research is as follows:1.In order to represent activities in event logs as real-valued vectors in a distributed semantic space,this thesis is aimed to design a novel methodology based on the difference between the event in business process and instances of the word in sentence.To address the limited-volume and sparsity issue of event logs,the raw log was enhanced through simulation of multiple process models.Furthermore,a vector representation learning algorithm was developed based on a negative sampling method,which has the capability to learn both activity vectors and temporal interval vectors from the enhanced event log.2.In handling the different types of ongoing process instances,a transfer learning framework for remaining time prediction is designed.The framework aims to construct multiple prediction models for trace in event log with different lengths,each of which is based on bi-directional recurrent neural networks with attention mechanisms.The context information of business process instances is modeled by the bi-directional recurrent neural networks,and the attention mechanism is used to assign different weights to each step3.Large-scale experimental studies are carried out to comprehensively evaluate the effectiveness of the temporal activity vector and the transfer learning framework for remaining time prediction.The experimental results show that the proposed method is significantly better than the existing process-model-based approaches and Recurrent-Neural-Network-based approaches.Meanwhile,the analysis of the framework suggested that the performance improved by the whole framework but not the part of it.