Similarity resonance for improving process model matching accuracy

Comparing and matching process models is a common task in many process engineering applications such as in querying and refactoring large process model repositories. The first and most challenging task in process model matching is to establish a similarity measure between process activities. Existing works use label similarity techniques to compute activities' similarity. In this paper, we propose a contextual similarity measure that, in addition to label similarity, exploits the similarity of the context surrounding activities. We introduce similarity resonance, a recursive algorithm that computes a global contextual similarity between process activities. Our intuition is that two activities are similar if they are executed in similar contexts which are represented by the surrounding activity neighbors. In turn, the activity neighbors are similar if their neighbors are similar, and so on. In this way, the pairwise similarity between all activities is iteratively computed and updated based on the similarity between their neighbors until the similarity scores have been stabilized and have been propagated to the whole graph. The approach has been implemented as a ProM plugin and was evaluated using several real-life datasets.