Multi-perspective Next Event Prediction in PPM via Heterogeneous Graph Neural Networks

Predictive Process Monitoring (PPM) is a subfield of Process Mining that mainly focuses on using standard machine learning (ML) and deep learning (DL) methods to learn from historical data and predict the future of an ongoing process execution based on its early stages. While different ML and DL methods have been extensively explored in the PPM literature, a new type of technique has emerged in recent years. These techniques involve Graph Neural Networks (GNNs), an innovative type of neural network that only a few approaches in PPM have applied so far. GNN models, and even more Heterogeneous GNN (HGNN), offer the advantage of working with a more natural representation of complex sequences such as execution traces, as well as dynamic systems such as process models, thus allowing for a more expressive and semantically rich encoding. This work presents SEPHIGRAPH, an approach that utilizes an HGNN model to process a graph encoding of the traces to tackle the multi-perspective next event prediction task, which involves predicting the full set of attributes of the next event to be performed. We evaluate its performance on multiple real-world datasets and compare it against other state-of-the-art approaches. The results show that SEPHIGRAPH is able to outperform existing approaches both in the next activity prediction task, as well as in the multi-perspective next event task.