Towards an enhanced next activity prediction using attention based neural networks

Predictive Business Process Monitoring has gained significant attention in recent years due to its capability to forecast process outcomes and predict the next event within an ongoing process. In the last few years, works have focused on deep learning and its applications in predicting the next activity. Some research used Long Term Short Memory, while others used Convolutional Neural Networks. However, long-term short-term memory models have the constraint of relatively slow training, while Convolutional Neural Networks are fast but may reach lower accuracy. In previous studies, researchers attempted to use meta-attributes to improve the accuracy of next activity prediction. However, in this research, a network that includes a layer of Bi-Directional Long Term Short Memory and utilizes attention mechanisms is introduced for an enhanced next activity prediction. Attention mechanisms in this research help determine the dependencies and importance of the activities within the process. At the same time, this network benefits from the fast training properties of the Convolutional Neural Network. This research also notes that gathering meta-data is not always efficient for businesses and may be overlooked in real-life systems. Therefore, only activity and sequence are used as input and managed to surpass the state-of-the-art deep learning models in designated experiments. Additionally, a case study was conducted on Order to Cash to test an Attention-based Multi-Channel Convolutional Neural Network in a real-life situation. This model achieved an accuracy of 87% in the case study.