Semantic Trace Topics: Text-Based Encodings for Interpretable Multi-Faceted Process Exploration

Trace encodings are central to process mining, as they determine what aspects of an event log can be explored and how results can be interpreted. Existing approaches to multi-faceted exploration often rely on high-dimensional boolean feature vectors over events, transitions, and attributes, which suffer from the curse of dimensionality and hinder semantic interpretation. We propose to represent event sequences as text documents and apply non-negative matrix factorization ({NMF}) topic modeling to obtain low-dimensional, interpretable trace features. A flexible, domain-specific vocabulary captures events, n-grams, and discretized event attributes, yielding topics that describe recurrent behavioral patterns as weighted sets of human-readable terms. Each trace is embedded as a topic-weight vector that supports classification, clustering, and dimensionality reduction. On a real-world traffic fines log, topic-based representations reveal meaningful process variants and their relations to outcomes such as success, duration, and cost.