Exploiting Temporal Convolution for Activity Prediction in Process Analytics

Process Mining (PM) is meant to extract knowledge on the behavior of business processes from historical log data. Lately, an increasing attention has been gained by the Predictive Process Monitor- ing, a field of PM that tries to extend process monitoring systems with prediction capabilities and, in particular. Several current proposals in literature to this problem rely on Deep Neural Networks, mostly on recurrent architectures (in particular LSTM) owing to their capability to capture the inherent sequential nature of process data. Very recently, however, an alternative solution based on a convolutional architecture (CNN) has been proposed in the literature, which was shown to achieve compelling results. Inspired by this line of research, we here propose a novel convolution-based deep learning approach to the prediction of the next activity which relies on: (i) extracting high-level features (at different levels of abstraction) through the computation of time-oriented dilated convolutions over traces, and (ii) exploiting residual-like connections to make the training of the predictive model more robust and faster. Preliminary results on real-life datasets confirm the validity of our proposal, compared with an LSTM-based and a CNN-based approach in the literature.