Audine Subias

Predictive process monitoring is at the intersection of machine learning and process mining. This subfield of process mining leverages historical data generated from process executions to make predictions about the ongoing process. One of the predictive process monitoring tasks with high interest is predicting the remaining cycle time of process instances. Recently, deep neural networks, particularly long short-term memory, have attracted much attention due to their ability to learn relevant features automatically and predict with high accuracy. While these models require data defined in the Euclidean space, graph neural networks have the advantage of handling data that can be represented as graphs. This paper proposes the use of graph neural network models to predict the remaining cycle time, which has not yet been studied in the literature. The proposed models are evaluated on real-life event logs and compared to a state-of-the-art long short-term memory model. The results show that graph neural network models can improve prediction accuracy, particularly for complex processes.

Checking the diagnosability of a timed discrete-event system usually consists in determining whether a single fault event can always be identified with certainty after a finite amount of time. The aim of this article is to extend this type of analysis to more complex behaviors, called event patterns, and to propose an effective method to check diagnosability with the use of model-checking techniques. To do so, we propose to convert the pattern diagnosability problem into checking a linear-time property over a specific time Petri net.

The overall context of this paper is the event-based behavior analysis and focuses on modeling and analyzing behaviors of interest involving time information. Any behavior of interest from any time event system is concisely defined as a set of time constrained events that must occur (positive behavior) and a set of time constrained events that must not occur (negative behavior). This article proposes a formal extension of the chronicle formalism that allows for the concise description of positive and negative behaviors. Based on this new formalism, several criteria are introduced, they formally characterize and compare a set of chronicles. A fully proved implementation of the proposed criteria is then described; it relies on the use of polyhedron techniques to solve systems of linear inequalities.