Process mining usually explains an event log with a single model, clustering, or classifier fitted to every case. The TracePrism paper, posted on arXiv, argues that real logs often mix cases following different sub-processes, so a one-size-fits-all analysis can be misleading. Trace clustering is the standard way to separate these sub-populations, and TracePrism builds on that idea.
TracePrism extends semantic trace topics to iterative, encoding-aware trace partitioning. Rather than a single static split, the method appears to refine partitions in steps while taking the encoding of traces into account. The abstract positions this as a way to better handle heterogeneous logs, though the full details of the algorithm and evaluation are not included in the abstract.