| In the current rapidly changing business environment,companies are constantly looking for the best way to improve business efficiency,so as to improve the company’s production efficiency.In order to better reflect actual business processes from a large amount of data and extract valuable information,a new automated technology,process mining,has emerged.Process mining technology focuses on extracting knowledge from business processes from event logs.However,in many cases,because event logs are not typically generated specifically for process mining applications,the event granularity of event logs may be too fine.Having event logs at an appropriate level of abstraction is crucial for successfully applying process discovery technology.The interaction between users and computers generates a large amount of interactive operation data,which can be captured and recorded as user interaction(UI)logs.Generally,the granularity of events in UI logs is very fine.If you want to apply process mining techniques to such logs and discover user behavior models,you must abstract UI events.This paper transforms the event abstraction problem into an event classification task,and constructs a mapping from low-level events to high-level events using neural network training and learning.LSTM(Long and Short Term Memory Neural Network)is a special cyclic neural network that can capture long-term dependencies.Based on this characteristic,we introduce Long and Short Term Memory Neural Network into event classification.The research in this article is mainly divided into the following parts:(1)In terms of obtaining interactive data,we have developed a tool-UI Logger-for capturing user and computer interaction data in the Python environment.The Pyhook3 module library is used to monitor mouse and keyboard events,and the Pywin32 module library is also used to obtain Windows window information.Use OCR character recognition technology to identify and obtain field names and field values entered by the keyboard.Finally,we obtained user interaction logs of students using software EA for software engineering modeling.(2)In the data pre-processing stage,the data set was coded using the One-hot coding technique,which is a common coding format for textual data and is a method of digitising features,and the pre-trained word vector file,which is also used in view of the limitations of One-hot coding in increasing data volume.Word embedding was performed for columns that could not be encoded with One-hot.(3)A long and short-term memory network(LSTM)was constructed for predictive classification of events,and the model was systematically evaluated using the metrics for multi-classification model evaluation: accuracy,precision,recall and F1 score.At the stage of constructing the network model,the basic LSTM network was applied,which had a predictive classification accuracy of 91%;to improve the classification of the model,we then designed a two-layer LSTM network on top of this,which improved the classification evaluation metrics compared to the basic LSTM network.Finally,the model tends to enter the phenomenon of overfitting,i.e.the accuracy of the training set is relatively high and the accuracy of the test set is relatively low,and adjusting the model parameters cannot cause a change in the accuracy.In this regard,we introduced the Dropout mechanism to improve the generalisation ability of the model.We have also achieved good results in comparison with other studies. |
