Neuro-symbolic predictive process monitoring for multi-attribute suffix prediction

Leveraging historical process execution data to predict the evolution of an ongoing process is a challenging task in Business Process Management (BPM). In recent years, Transformer-based models have shown promising results in multi-attribute sequence prediction, which refers to the task of forecasting not only the next activity in a process but also its associated attributes (e.g., timestamp, resource, and data values). However, while the accuracy of these models is generally high, there are specific situations where background knowledge about the process can be beneficial. In particular, such knowledge can enhance prediction quality for outlier executions in the training data or when process changes occur due to concept drift in the test data. In this paper, we propose a principled approach to integrating prior knowledge into an autoregressive Transformer-based model for multi-attribute sequence prediction at test time. This knowledge is expressed using Multi-Perspective Declare, a process modeling language capable of describing process behavior from various perspectives, including control-flow (i.e., the expected sequence of activities to be executed in the process), data-flow (i.e., the data attributes associated with these activities during process execution), resources, and timestamps. We evaluate our approach using real-life logs and demonstrate that incorporating background knowledge consistently enhances the performance of multi-attribute sequence prediction, especially in rare and exceptional situations.