Learning context-based representations of events in complex processes

Abstract

Process mining techniques enable the understanding and optimisation of business processes through the analysis of event logs. In this area, it is essential to understand the relationships between the executed activities and their attributes, i.e. their execution context, especially in problems such as predictive monitoring when deep learning techniques are applied. However, the lack of robust methods to represent this execution context has been highlighted in the existing literature. This paper presents a new technique for generating event representations that can capture their execution context. Our approach, based on an autoencoder, allows pre-training of embeddings at activity level or any other categorical attribute by reconstructing context windows, providing contextual information about previous and subsequent events. Experiments on 21 datasets of real-world processes show how the embeddings generated by our approach can improve the accuracy of three state-of-the-art prediction models in the task of predicting the next activity. The results, which are particularly remarkable in the case of complex processes, demonstrate the importance of the capture of contextual information for better predictions.