Distributional similarity between activities in certain and uncertain event data

Process mining provides a broad portfolio of methods for extracting insights from event data. Many of these methods can benefit from incorporating semantic relations between activities. A common way to capture such relations is distributional similarity, which links activities based on the contexts in which they occur. Distributional information can be encoded in activity embeddings learned either by prediction-based approaches (adapting neural models from natural language processing) or by count-based models. In prior work, we showed that simple count-based embeddings can match or even outperform neural prediction-based methods for assessing activity similarity in event data. This finding, together with the limited interpretability and high computational cost of neural approaches, motivates the use of count-based alternatives. However, existing distributional similarity methods are limited to certain event data, where each event is assigned a single activity label. In this paper, we argue that many contemporary process mining applications also involve uncertain event data, where events are associated with multiple possible activities and corresponding likelihoods. We generalize count-based approaches and adapt state-of-the-art neural network approaches to the domain of uncertain event data. For both count-based and neural network approaches, probabilistic evidence is incorporated to provide a more robust and comprehensive similarity assessment. Finally, we contribute a comprehensive benchmarking framework for distributional similarity between activities on both certain and uncertain event data, evaluating (i) intrinsic similarity quality, (ii) performance in next-activity prediction as a downstream task, and (iii) computational efficiency. Across 28 certain and 9 uncertain event logs, count-based activity-activity embeddings achieve the highest intrinsic similarity quality overall and allow a simple activity-only predictor to match the state-of-the-art Tax model in next-activity prediction without temporal features, while running more than an order of magnitude faster than neural alternatives on uncertain data. Embedding quality is task-dependent: activity-activity representations lead in intrinsic retrieval while activity-context representations lead in downstream prediction.