Improving the non‐compensatory trace‐clustering decision process. (23rd September 2021)

• Record Type: Journal Article
• Authors: Delias, Pavlos; Doumpos, Michael; Grigoroudis, Evangelos; Matsatsinis, Nikolaos
• Other Names: Moreno‐Jiménez José María guestEditor.; Vargas Luis G guestEditor.
• Is Part Of: International transactions in operational research. Volume 30:Number 3(2023)
• Page Start: 1387
• Page End: 1406
• Publication Date: 2021-09-23
• Subjects: process mining -- trace clustering -- outranking methods -- robustness; Operations research -- Periodicals; 003
• DOI: 10.1111/itor.13062
• Languages: English
• ISSNs: 0969-6016

Abstract

Abstract: In flexible environments (such as healthcare or customer service), the observed behavior is expected to considerably vary, namely there is no dominant flow path. Such a high variability obstructs the process discovery task since it regularly leads to "spaghetti" process models. Trace clustering is about grouping behaviors, and discovering a distinct model per group, thus delivering more comprehensible results. In previous works, we have proposed a multiple‐criteria non‐compensatory approach to create a similarity metric and finally perform trace clustering. The main problem that we tried to respond to is how to summarize a process event log, when a lot of variability exists, thus facilitating knowledge discovery. The underpinnings of the non‐compensatory approach are first the fact that a sufficient number of criteria must be concordant with the similarity (concordance setting) and second that there should not exist any criterion raising a veto logic, that is, among the criteria that are not concordant, none of them must be conflicting with the similarity (discordance setting). This work challenges improved support for the decision‐maker (DM) and it extends the previous approach by (i) proposing an improved clustering technique based on spectral clustering; (ii) guiding the clustering process by allowing reinforced or counterveto effects and pairwise constraints; (iii) handling outliers through a trimming approach as an integer linear program. All improvements aiming at making elements of the trace‐clustering process more accessible to the DMs and enhancing the understandability of the analysis.