| Process mining technolog can extract interested knowledge from sufficient historical event logs and provide greater business value for process stakeholders.As important research direction of process mining,predictive process monitoring technology is committed to the analysis of currently running process instances,including remaining time prediction,next activity prediction and result-oriented prediction.Making full use of the behavior relationship or graph information in the log will improve the prediction ability of the prediction model.The main research contents of this paper are as follows:(1)As a preprocessing task of result prediction,an event log prediction and repair method based on Convolutional Neural Networks(CNNs)is proposed,aimed at solving the problem of incomplete historical logs caused by missing information.The core idea is that according to both the behavior relationship between activities and the two dimensions attributes,that are time attribute and activity attribute,the event log of business process is converted into spatial data,which is then converted into image matrix,and the missing activities are predicted through the training of convolutional neural network model.The proposed method does not depend on any prior knowledge about the business process of generating event logs.Finally,real and artificial event logs are generated to compare the proposed method with the existing research results.Experimental results show that the proposed method is superior to the existing research results in the accuracy of activity repair.(2)Random forest(RF)and XGB(e Xtreme Gradient Boosting),as tree-based ensemble learning methods,are considered to be efficient machine learning methods in predicting next tasks or outcomes.However,there exist the problems of low variance,high deviation and high variance,low deviation between the prediction results and the actual values of random forests and XGB,respectively.In order to overcome the two shortcomings,which are the variances and deviations of the prediction results,actual values of the two prediction methods cannot be reduced at the same time,this paper puts forward a method LCE(Local Cascade Ensemble)that applies XGB building block classifier and uses Bagging(Bootstrap AGgregation)to reduce variances.In order to solve the problem that the location relation of the activity is lost in the process of log conversion code,an additional sequential and concurrent behavior is proposed,and the neighbor activity of the target activity is considered at the same time.Experiments are carried out on seven common datasets,and the experimental results show that the proposed prediction method is more accurate than the benchmark method,and the use of log behavior information is significant to improve the accuracy of the prediction method.(3)The graph information in the log is widely used in the result prediction task,but no work has been done to make a comparative study of the use of graph information.We compare four machine learning methods based on graph features and two graph neural networks(GNNs),then use LSTM(Long Short-Term Memory)and bidirectional Long Short-Term Memory Network(Att-BLSTM)with attention mechanism as strong baselines.The use of graph information plays an extremely important role in improving the quality of result prediction.Experimental results show that simple machine learning methods based on graph features are generally superior to deep learning models in result prediction tasks,and the methods based on graph features definitely have higher prediction precision and accuracy.Figure[15] Table [11] Reference [99]... |
