Learning Context-Based Representations of Events in Complex Processes

Process mining techniques need to know not only the activity that is recorded at each moment but also its execution context, i.e., how the activity is related to the rest of process activities. This information is very relevant for addressing many process mining problems, such as predictive monitoring, especially when applying deep learning techniques. However, the representation of this execution context has not received much attention in the literature. In this paper, we present a novel approach based on an autoencoder for the generation and training of embeddings that properly capture the contextual information related to the execution of an activity. This deep learning-based architecture allows to obtain representations at the level of individual activity, roles, or any other categorical attributes. To evaluate this proposal, an experimentation has been conducted that shows how the embeddings generated with the proposed autoencoder allow to improve the accuracy of the next activity prediction with three of the main state-of-the-art predictive models. Particularly noteworthy are the results on complex processes, where the contextual information provided by our embeddings can significantly improve the state-of-the-art results.