A global attention-based convolutional neural network for process prediction

Process prediction, which analyzes process information in event logs to predict upcoming activities and their duration, has become one of the main drivers for moving the process mining field forward. However, the unbalanced distribution of data in event logs and the neglect of event trajectory correlations affect the final prediction accuracy. In this paper, a novel convolutional neural network is proposed to predict the upcoming events using prefix event visualization. Firstly, a process global focus block that is more suitable for the distribution of event log image data is proposed, thus focusing more on the process relevance information contained in the images. Based on this block, we propose a process global attention network to predict the next activity of an ongoing event, effectively improving the performance associated with the process prediction algorithm. Secondly, the training of the process global attention network is optimized by introducing a multi-class focal loss function, thus reducing the effect of imbalance in the event log samples. It is experimentally verified that the proposed process prediction network improves the global feature relevance and achieves progressive results in terms of prediction performance through effective training.