Process mining aims to extract information from recorded process data, which can be used to gain insights into the process. This requires applying a discovery algorithm and settings its parameters, after which the discovered process model should be evaluated. Both steps may need to be repeated several times until a satisfying model is found; we refer to this as process exploration. Existing commercial tools usually do not provide models having executable semantics, thereby disallowing for accurate map evaluation, while most academic tools lack features and by the repetitive nature of process exploration, their use is tedious. In this paper, we describe a novel process exploration tool: the Inductive visual Miner. It aims to bridge this gap between commercial and academic tools, by combining the executable semantics of academic tools with the exploration support of commercial tools. It also adds animation and deviation visualisation capabilities.
To gain insights in business processes based on factual knowledge, recorded event
data can be analysed using process mining. Process mining aims to extract
information from recorded process data, stored in an event log, and starts with
discovering a process model from the event log. However, many process
discovery algorithms exist, their parameters have to be set, and the question at hand
might require to focus on speci c parts of the event log. The implications of these
choices are, although well-studied for academic approaches, unclear for the
average user, which makes it di cult to obtain a model that is suitable to answer
the question at hand. In this paper we focus on process exploration, which is the
process of repeatedly trying settings until a satisfactory model is discovered [
The rst step to take in process exploration is to select a process discovery algorithm and to set its parameters. Moreover, the scope of the exploration needs to be set by applying all kinds of lters and choosing a perspective, e.g. one can focus on the control ow or resource perspective.
Copyright c 2014 for this paper by its authors. Copying permitted for private and academic purposes.
In the second step of the exploration cycle, one needs to apply the algorithm in the selected scope to the event log to ob- set scope tain a process model. Before conclusions use can be drawn and insights can be gained, process map evaluate prodcisecsosvmerap the model should be evaluated. For in- process map stance, compliance related questions, such Figure 1: Exploration cycle. as whether the four-eyes principle was adhered to, can only be answered if the model represents a large part of the behaviour in the event log, and future related questions should only be answered using models that are likely able to represent future behaviour. Evaluation of a model with respect to an event log can only be done accurately if the behaviour that the model allows is well-de ned, i.e. if it has executable semantics, and di erent parts of the model might have di erent problems.
Often, general questions, such as what a process looks like, lead to more speci c questions such as where in the process delays or deviations occur, or to questions that need to be answered using other perspectives on the event log. Or, the evaluation shows that the question cannot be answered with the discovered process model. Then, the parameters need to be set again and a new model must be discovered; process exploration is a highly iterative process.
After a user has found a suitable model, that model can be used in for instance
automatic enactment of models in systems [
Current commercially available process exploration tools o er plenty of options to set the scope of the exploration, but usually do not produce models having executable semantics, which thus cannot be used for automated evaluation or further use. There is plethora of academic tools available to set the scope of the exploration, to discover a process model and to evaluate it, but given the nature of process exploration, using them iteratively is tedious. In this paper, we introduce a tool, Inductive visual Miner (IvM), that aims to bridge this gap between commercial and academic tools. It supports the steps of process exploration by chaining existing academic tools and streamlining their use. Moreover, it improves on evaluation by a new notation and the addition of animation and quick node selection ltering. Thus far, such capabilities only existed for tools having no or just weak semantics or without formal guarantees (Fuzzy Miner, Disco, BPMjOne, Celonis, Perceptive, etc.).
IvM has been implemented as a plug-in of the ProM framework, which can be obtained by installing ProM 6.4 from http://promtools.org and, using the ProM package manager, installing the plug-in Inductive visual Miner. Example event logs can be obtained from http://www.processmining.org/ logs/start; a screencast is available at http://vimeo.com/user29103154/ inductivevisualminer.
In the remainder of this paper, we explain the implementation of IvM,
highlight the deviation visualisation and give an example. For a detailed comparison
with existing exploration approaches, please refer to [
Inductive visual Miner: Implementation The architecture of IvM resembles a chain of analysis and visualisation tasks, shown in Figure 2. To encourage exploration, a user can change any parameter at any time. IvM will ensure that the current computation is discarded and the chain is restarted from the rst task that is in uenced by the parameter change. For instance, if the user selects or deselects a node, only the tasks ` lter node selection' and `animate' are redone. As especially the align task can take some time, intermediate visual results are shown to the user until the next task is nished.
model enriched model highlighted model animated model prepare log
lter activities discover align perspective activity threshold noise threshold lter node selection selected nodes animate
In the prepare log task, the events in the log are classi ed using the provided
perspective classi er. Next, in the lter activities task, given a threshold value,
the most-frequent activities are kept, the events of other activities are ltered
out. The Inductive Miner - infrequent (IMi) [
Once the model is available, it can be exported to ProM for further analysis,
both as a Petri net and as a process tree; a user can perform its own evaluation
without waiting for the evaluation of IvM to nish. At any point during the
exploration, the model can be saved as bitmap (png) and vector (pdf, svg) image
formats. The full animation of the complete log can be exported to bitmap (avi)
and vector (svg) based movie formats once it is computed.
1 At time of writing, we limited the quick preview to 50 traces for performance reasons.
Deviations. Deviations are a crucial part
of the evaluation: they show precisely C
what parts of the model deviate with 7 1
respect to the log. Deviations are
visualised to show shich parts of the model t 1
well and which parts do not. This is
important for drawing reliable conclusions. Figure 3: Model with the result of
Two types of deviations have been identi- the align task. The edge
circumvented [
(a) Default parameters.
(b) After a few iterations; with deviations and animation. In this paper, we discussed the cycle of process exploration, consisting of repeatedly setting parameters, discovering a process model and evaluating it. We identi ed a gap between existing commercial and academic process exploration tools: commercial tools usually do not provide models having executable semantics, thereby disallowing for accurate map evaluation, while most academic tools lack features such as seamless zooming and animation, thus do not support the repetitive nature of process exploration well.
We introduced a process exploration tool, Inductive visual Miner (IvM), that aims to bridge this gap. When started, IvM immediately applies a chain of analysis and visualisation tasks to show the user not only a model, but also the traces of the event log animated on it, and where the log and model deviate from one another. IvM encourages the user to interact by enabling setting parameters at anytime: computations will be restarted as necessary in the background. IvM is not as feature-rich as some of the commercial tools, but shows that it is possible to use powerful techniques with formal guarantees in a user-friendly package. We hope that IvM will inspire commercial vendors to consider models with executable semantics and support deviation analysis. Extensions to IvM can be made in all tasks, for instance other process tree discovery algorithms can be plugged in instead of IMi.
In the future, we'd like to include approximation algorithms to compute the alignments in order to speed it up. To allow for even better evaluation, several extensions are possible, such as global quality measures ( tness, precision and generalisation) and identi cation of traces in the animation. Furthermore, several other lters such as lters on speci c activity, timestamp, resource and on data could be included to give a user more freedom in setting the scope.