act2vec, trace2vec, log2vec, and model2vec: Representation Learning for Business Processes

In process mining, the challenge is typically to turn raw event data into meaningful models, insights, or actions. One of the key problems of a data-driven analysis of processes, is the high dimensionality of the data. In this paper, we address this problem by developing representation learning techniques for business processes. More specifically, the representation learning paradigm is applied to activities, traces, logs, and models in order to learn highly informative but low-dimensional vectors, often referred to as embeddings, based on a neural network architecture. Subsequently, these vectors can be used for automated inference tasks such as trace clustering, process comparison, predictive process monitoring, anomaly detection, etc. Accordingly, the main contribution of this paper is the proposal of representation learning architectures at the level of activities, traces, logs, and models that can produce a distributed representation of these objects and a thorough analysis of potential applications. In an experimental evaluation, we show the power of such derived representations in the context of trace clustering and process model comparison.