Process mining techniques are meant to extract non-trivial information from complex data. Controlled experiments of the algorithms underlying process mining techniques often require logs of process executions that fit the specific purposes of each specific test. Therefore, many tools for the log generation from both procedural models (e.g., Petri nets or BPMN models) and declarative models (e.g., based on LTL or Declare) have been developed. However, the log generation from declarative models still lacks tools for log generation that address specific purposes such as the specification of trace length distributions, the setting of the number of variants that should appear in the log, or the specification of the number of activations of a constraint that should be contained in a trace. We address this research gap by proposing an extension of the Declare4Py Python library that generates synthetic event logs using an Answer Set Programming-based solution whose eflxibility supports the encoding of specific purposes.
Process Mining (PM) is a research area that analyzes the execution data of a business process
(an event log) to extract useful information for process improvement. This is not a trivial task as
event logs are sets of process instances (a.k.a. traces) that are a complex type of data. A trace is
composed of a sequence of events arranged in chronological order and each event contains a set
of attributes that can be both symbolic (e.g., the name of the activity executed or the involved
resource) and numeric (e.g., a timestamp). In addition, process instances are samples of a process
model, that is, background knowledge that constraints the execution order of the events. This
background knowledge can be expressed with procedural models (e.g., Petri nets or BPMN
models), which specify the exact control flow of the activities in a trace, or with declarative
models (that is, constraints expressed in Linear Temporal Logic on finite traces or Declare [
The controlled evaluation of PM algorithms requires event logs that fit the purposes for
which each specific experiment has been designed. For example, one purpose can be to test
how the performance of an algorithm is afected when the distribution of trace lengths in the
event log varies. However, although several tools have been developed for generating event
logs by simulating (declarative and procedural) process models [
The ASP-based solution adopted in Declare4Py takes as inputs the number of traces to generate, the minimum length and the maximum length of a trace (i.e., the number of events in the trace), a process model containing MP-Declare constraints, and encodes these inputs in an ASP program to be solved by the Clingo solver (https://potassco.org/). The solution is a set of traces that satisfy the input model and the minimum/maximum length specification. We leverage this solution to generate event logs with the following four purposes as additional inputs.
Users can specify a trace length distribution, that is, an input probability distribution on the lengths of the traces in the log. Declare4Py supports three probability distributions: i) the custom distribution, where the user has to specify the probability of a generated trace to have length for each ∈ [, + 1, . . . , ]; ii) the uniform distribution where all the trace lengths have the same probability to appear, i.e., 1/( − + 1); iii) the Gaussian distribution where the user has to specify a mean and a variance for sampling the trace lengths from a Gaussian distribution. Once has been sampled, the Clingo solver is called − + 1 times for each trace length .
Users can specify the number of variants , so the generated event log can be segmented into groups of traces having the same control-flow (i.e., the same sequence of activity names) but that can difer for other attributes, such as the timestamps or the resources. Here, Clingo is called diferent times: in the first call, traces are generated representing the variants; then, for each variant, Clingo is called to generate traces with the control-flow of the variant. In this second call, the input ASP program also encodes the control-flow of the variant.
Users can specify a subset of constraints in the input model to generate negative traces, that is, traces that do not satisfy (at least one of/all) the constraints in the specified subset. This allows users to obtain event logs with labelings defined by MP-Declare constraints. Such labeled traces can be used to train and test Machine Learning-based process mining algorithms. The positive traces satisfy all constraints in the input model, whereas the negative traces do not satisfy the specified subset of constraints. Also in this case, the subset of constraints to be violated is encoded in an ASP program.
Users can specify the number of activations for a subset of constraints in the input model. For example, the user can specify for the constraint Response[(CRP, J), ReleaseB] an activation number of 3 that means that the event with activity name CRP with payload J should occur three times. The user provides as input a range of possible values for the number of activations for a specific constraint and then Declare4Py randomly selects a value in that range for each generated trace. The subset of the constraints for which the number of activations is specified and the user-defined value range for the number of activations are encoded in ASP.
The better computational times of the adopted ASP solution with respect to the
state-of-theart (declarative) log generators have already been shown in [
The GitHub repository of Declare4Py4 contains the source code of the tool and all the tutorials. A specific tutorial for the ASP-based log generator 5 shows how to run the log generator using all the available options. The video presentation of this paper can be accessed at https://www.dropbox.com/scl/fi/cbihgbw34smkisb7u1sry/ Screen-Recording-2023-09-12-at-19.17.17.mov?rlkey=pvzt6cuj5yk98azi611zd79xy&dl=0. 4https://github.com/ivanDonadello/Declare4Py 5https://github.com/ivanDonadello/Declare4Py/blob/main/docs/source/tutorials/9.Log_Generation.ipynb