Using Convolutional Neural Networks To Solve The Problem Of Predicting The Next Process Event

Adding the feature of business process event prediction to information systems(IS)increases their productivity in the long run,enhances the quality of the taken decisions,and eliminates system anomalies.An event is that step in any information system that the shareholders need to know beforehand.So,they can take decisively correct decisions on the consequences of that event.The recent works that focused on solving the problem of predicting the next process event used different approaches.The previously used methods to tackle this problem were complicated,and that badly affected the training speed.Also,the methods that the last studies utilized did not yield highly accurate predictions on all the datasets they used to train and validate/test their models.Inspired by the previous works of employing deep learning approaches to predict the next process event based on files of logged events,we propose the use of one-dimensional convolutional neural networks(1D CNN)in two architectures to address the same problem.In the proposed approaches,we significantly enhanced the results of the state-of-the-art studies on all the provided eight datasets.We also enrich the literature body of using 1D CNNs in an NLP-like problem which is rarely used as a CNN architecture.Moreover,we comparatively analyzed the previous studies and the proposed methods to show the significance of our results.The main contents of this thesis can be summarized as follows:1)Predicting the next business process event using convolutional neural networks.In this study,we used a five-layer 1D CNN method to predict the next process event based on previous instances.This approach generated significantly high predictions.Moreover,this method demonstrated high training speed that it processed more than 8000 words per second.2)Predicting the next business process event using Res Nets.In this study,we created six Res Net models based on the depth of the model.We first evaluated our proposed Res Net by modifying one of its important hyperparameters specifically,the number of residual blocks.After that,we compared the achieved precision,recall,and accuracy of the predictions of the resulted Res Nets with our proposed Res Net.That comparison made it clear,adding more residual blocks to our proposed method will not add any significant improvements to the overall performance.However,this method outperformed the first one that used a five-layer 1D CNN.3)We conducted a comparative analysis between our proposed approaches and the state-of-the-art studies.Our proposed approaches outperformed the state-of-the-art methods on all the datasets across all the performance metrics rates.The statistical comparisons that we conducted showed a significant difference between our performance and the previous studies.Also,we compared our two proposed methods to each other and found that using 1D CNNs in a Res Net architecture produces more accurate,precise with high recall rates predictions than plain 1D CNNs.4)A framework that lives on a cloud server and supplies RESTful links is proposed.The supplied links can be consumed using any system in order to utilize the power of our proposed deep learning algorithms.