A Knowledge-based abstraction framework for trace comparison and semantic process mining

Event logs constitute a rich source of information for several process analysis activities, which can take advantage of similar traces retrieval. The capability of relating semantic structures such as taxonomies to actions in the traces can enable trace comparison to work at different levels of abstraction and, therefore, to mask irrelevant details, and make the identfication of similar traces much more flexible. For this reason, this thesis proposes a trace abstraction mechanism based on domain knowledge, which maps actions in the log traces to instances of ground concepts in an ontology, and then allows one to generalize them up to the desired level. Abstracted traces are also provided as an input to semantic process mining; finally, abstracted models (i.e., models mined from abstracted traces) can be compared and ranked, by adopting a similarity metric, able to take into account penalties collected during the abstraction phase. The overall framework has been tested in the eld of stroke management, where we were able to cluster similar traces, corresponding to correct medical behaviors, abstracting from details, but still preserving the capabilities of identifying outlying situations. Moreover, we could mine process models that are easier to interpret, since unnecessary details are hidden, but key behaviors are clearly visible. Finally, we were able to rank abstract process models more similarly to the ordering provided by a domain expert, with respect to what could be obtained when working on non-abstract ones.