Many commercial information systems and enterprise resource planning tools
routinely adopted by organizations and companies worldwide, like those provided
by, e.g., Oracle and SAP, record information about the executed business process
instances in the form of an event log [
Event logs constitute a very rich source of information for several business
process management tasks. Indeed, the experiential knowledge embedded in
traces is directly resorted to, e.g., in operational support and in agile workflow
tools, which can take advantage of trace comparison and retrieval.
Operational support [
In my PhD thesis, I am developing a framework to compare and retrieve process traces, represented at di↵erent levels of abstraction . The framework will then be interfaced to operational support or agile workflow tools, as well as to other analysis mechanisms. In this paper, I describe the methodological approach behind trace abstraction; the applications mentioned above will be considered in my future work.
Copyright © 2017 for this paper by its authors. Copying permitted for private and academic purpose. In Proceedings of the ICCBR 2017 Workshops. Trondheim, Norway
1.1
We are developing a semantic-based, multi-level abstraction mechanism, able to operate on event log traces. In our approach, actions in the log are mapped to instances of ground concepts (leaves) in a taxonomy, so that they can be converted into higher-level concepts by navigating the hierarchy, up to the desired level, on the basis of the user needs.
The abstraction mechanism has been designed to properly tackle nontrivial issues that could emerge. Specifically: – two actions having the same ancestor in the taxonomy (at the chosen abstraction level) may be separated in the trace by a delay (i.e., a time interval where no action takes place), or by actions that descend from a di↵erent ancestor (interleaved actions henceforth). Our approach allows to deal with these situations, by creating a single macro-action, i.e., an abstract action that covers the whole time span of the two actions at hand, and is labeled as the common ancestor; the macro-action is however built only if the total delay length, or the total number/length of interleaved actions, do not overcome proper admissibility thresholds set by the user. The delays and interleaved actions are quantified and recorded, for possible use in further analyses. In particular, we have defined a similarity metric where this information is accounted for as a penalty, and a↵ects the similarity value in abstract trace comparison; – abstraction may generate di↵erent types of temporal constraints between pairs of macro-actions; specifically, given the possible presence of interleaved actions, we can obtain an abstracted trace with two (or more) overlapping or concurrent macro-actions. Our approach allows to represent (and exploit) this information, by properly maintaining both quantitative and qualitative temporal constraints in abstracted traces. Once again, this temporal information can be exploited in further analyses. In particular, the similarity metric we adopt in trace comparison can deal with all types of temporal constraints.
Specifically, the procedure to abstract a trace operates as follows: – for every action i in the trace: • i is abstracted as its ancestor at the taxonomy level selected by the user; the macro-action m i, labeled as the identified ancestor, is created; • for every element j following i in the trace: ⇤ if j is a delay, its length is added to a variable tot delay, that stores the total delay duration accumulated so far during the creation of m i; ⇤ if j is an interleaved action, its length is added to a variable tot inter, that stores the total interleaved actions durations accumulated so far during the creation of m i; ⇤ if j is an action that, according to domain knowledge, abstracts as the same ancestor as i, m i is extended to include j, provided that 1. RESEARCH SUMMARY tot delay and tot inter do not exceed domain-defined thresholds. j is then removed from the actions in the trace that could start a new macro-action, since it has already been incorporated into an existing one; • the macro-action m i is appended to the output abstracted trace which, in the end, will contain the list of all the macro-actions that have been created by the procedure.
The variables tot delay and tot inter, accumulated during abstraction, are also provided as an output attribute of each macro-action and they will be used as a penalty in abstracted trace similarity calculation.
The most significant and original methodological contributions of the work
thus consist in:
1. having defined a proper mechanism for abstracting event log traces,
able to manage non trivial situations (originating from the treatment of
interleaving actions or delays between two actions sharing the same ancestor);
2. having provided a trace comparison facility, which resorts to a
similarity metric (extending the metric presented in [
In the third year, I will concentrate on experimental work referring to trace comparison and I will deal with operational support, agile workflow management, or other activities, including process mining on abstracted traces. As regards process mining, in particular, we wish to test when abstraction allows to make clear and more readable process model. 1.2
With the help of an expert physician in stroke patient management, we have
formalized medical domain knowledge in a taxonomy (which has been organized
by goals) by using the Prot`eg`e ontology editor [
The system architecture we have developed, is shown in Figure 1. Rectangles represent computational modules, while ovals and cylinders represent domain knowledge sources and the database. The first step to be executed is event log preparation, that takes in input the available database (DB), and exploits domain knowledge (the taxonomy); the event log will then undergo abstraction. The abstracted event log will be given as an input to trace comparison resorting to the metric we have developed, or to process mining, operational support, or other activities, that we plan to realize by resorting to ProM.
Event log preparation
Eventlog
Multi-level abstraction
Abstracted event log
Trace comparison Process mining Operational support Agile workflow management
Taxonomy During my last PhD year, the framework will be tested in the field of stroke management, where we will adopt multi-level abstraction and trace comparison to cluster event logs of di↵erent stroke units, in order to highlight correct and incorrect behaviors, abstracting from details (such as local resource constraints or local protocols). The goal will be to show that, the application of the abstraction mechanism allows to obtain more homogeneous and compact clusters (i.e., able to aggregate closer examples), still making outliers clearly identifiable, and isolated in the cluster hierarchy. Some first encouraging results are already available.
As regards process mining, the ground processes (process learned on trace at the same level of taxonomy leaves) are typically ”spaghetti-like”: they presents an extremely large number of nodes and edges which make it hard to identify details. Our hypothesis is that models learned on abstracted traces will be much more compact and it will be possible for medical experts to analyze them. This topic will be studied during my last year as well.
Finally, we will provide abstracted traces as an input to operational support or agile workflow management facilities.