Research On Business Process Next Event Prediction Based On Deep Learning

The business process next event prediction task is a research focus in the field of predictive process monitoring,which can help practitioners allocate resources to activities in advance,prevent risky events,and improve decision making capabilities.However,as the volume and quality of log data increase,traditional machine learning methods are difficult to meet the requirements of the task,and deep learning methods provide a new direction to address this task.However,existing methods for applying deep learning still have limitations.On the one hand,researchers only focus on the temporal correlation between events,ignoring that logs can be abstracted into non-Euclidean graph structures,resulting in models with insufficient extracted feature information and prediction accuracy to be further improved.On the other hand,deep learning-based prediction methods require high data volume,and existing studies tend to improve only the network structure and ignore the interference of redundant information inside the logs,while such methods are difficult to meet industrial application scenarios due to the limitation of hardware devices.To solve the above problems,this paper proposes deep learning methods and strategies applicable to the task of process next event prediction,and the main research contents are as follows.(1)A process next event prediction method based on spatio-temporal feature fusion is proposed.First,the graph structure of each trajectory in the event log is modeled by establishing inter-node relationships,and a spatial feature extractor with a two-layer graph convolutional network is designed to process graphs with topological relationships;then,a temporal feature extractor with a three-layer ordered neuron long-and short-term memory network is constructed to obtain the hierarchical structure of higher abstraction levels between events,by fully considering the time-stamped properties of each event.solve the long-term dependency problem of long trajectories;finally,the two are combined to achieve fused spatio-temporal feature information and trained using a pre-defined iterative training approach.The comparison experiments are conducted in six public event logs,and the experimental results show that the prediction accuracy of the method is improved in each dataset,up to 8.63 percentage points.Meanwhile,the experimental results verify the necessity of making full use of inter-event correlation and learning the hierarchical structure of event logs to accomplish the prediction task,and the proposed method has good prediction accuracy and robustness.(2)A log sampling-based process next event prediction method is proposed,and the effects of different sampling methods and sampling rates on model prediction performance are evaluated.First,a pre-sampling strategy is proposed to update the event logs by log sampling technique to obtain high-quality and more representative logs to solve the problem of high time consumption caused by large data volume;then,a logrank++sampling method based on ranking is introduced to calculate activity importance,direct-following activity relationship importance,average activity importance of each trajectory and average direct-following activity Finally,a benchmark network is designed for the task of this paper,and a comparison experiment is conducted with the sampling strategy.The results show that the strategy effectively improves the prediction efficiency and accuracy,and the Logrank++ method significantly outperforms other methods,with a maximum efficiency improvement of 22.58 times at a preset sampling rate of 10% and a maximum accuracy of 99.94% at a preset sampling rate of 30%.In addition,combining the strategy with the spatio-temporal fusion network,the results show that the accuracy and efficiency of the combined method are significantly improved in each log,which can be improved by 18.9 percentage points and 41.41 times,respectively,proving that the strategy has good robustness and generalization ability.