In Process Mining, often one is not only interested in learning process models but also in answering questions such as “What do the cases that are late have in common?”, “What characterizes the workers that skip this check activity?” and “Do people work faster if they have more work?”. Such questions can be answered by combining process mining with classification (e.g., decision tree analysis). Several authors have proposed ad-hoc solutions for specific questions, e.g., there is work on predicting the remaining processing time and recommending activities to minimize particular risks. This paper reports on a tool, implemented as plug-in for ProM, that unifies these ideas and provide a general framework for deriving and correlating process characteristics. To demonstrate the maturity of the tool, we show the steps with the tool to answer one correlation question related to a health-care process. The answer to a second question is shown in the screencast accompanying this paper.
Introduction Process mining is not only about automatically learning process models. It also concerns with replaying event logs on the model to, e.g., check conformance or to uncover bottlenecks in the process. However, such analyses are often only the starting point for providing initial insights. When discovering a bottleneck or frequent deviation, one would like to understand why it exists. This requires the correlation of different process characteristics. These characteristics can be based on the control-flow (e.g., the next activity going to be performed), the data-flow (e.g., the amount of money involved), the time perspective (e.g., the activity duration or the remaining time to the end of the process), the organization perspective (e.g., the resource going to perform a particular activity), or, in case a normative process model exists, the conformance perspective (e.g., the skipping of a mandatory activity).
The study of these characteristics and how they influence each other is of crucial
importance when an organization aims to improve and redesign its own processes. Many
authors have proposed techniques to relate specific characteristics in an ad-hoc manner,
such as to predict the remaining processing time of a case or to analyze routing
decisions in the process or possible risks (see [
Copyright c 2014 for this paper by its authors. Copying permitted for private and academic purposes.
general correlation problem. The tool unifies the ad-hoc approaches described in
literature by providing a generic way to relate any characteristic (dependent variable) to
other characteristics (independent variables). Readers are referred to [
Starting point is an event log. For each process instance (i.e., case), there is a trace, i.e., a sequence of events. Events are associated with different characteristics, represented a key-value pairs. Mandatory characteristics are activity and timestamp. Other typical characteristics are the resource used to perform the activity, transactional information (start, complete, suspend, resume, etc.), and costs. However, many more characteristics can be associated to an activity (e.g., the age of a patient or size of an order).
The tool builds a table where each row corresponds to a different event and each column is a different characteristic. One of the columns become the dependent characteristic and the others are the independent characteristics; the relation between dependent and independent characteristics is discovered using decision-tree learning techniques. Before discovering the tree, the tool also allows some rows to be filtered out. For instance, one may want to only retain those events that refer to certain activities.
If a certain characteristic is valuable for an analysis but not present, our tool also allows extending event logs with additional characteristics that are not readily available. For instance, events can be extended with the remaining flow time till the end of the process instance or, also, the elapsed time since the process instance started. Other characteristics that may be added could be related to the resource who triggered an event (e.g., workload of the resource), i.e. who executed the respective activity. We can also add the next activity as a characteristic of an event. One can even add conformance checking results and external context information, such as weather information, to events as characteristics. In many cases, the values of these characteristics can be simply derived from the event log itself; in other cases, they need to be harvested from information sources outside the event log (weather information, stock index, etc.). Implementation. The tool is implemented as a plug-in of ProM, an open-source “pluggable” framework for the implementation of process mining tools in a standardised environment (see http://www.promtools.org). The ProM framework is based on the concept of packages each of which is an aggregation of several plug-ins that are conceptually related. Our new plug-in is available in a new package named FeaturePrediction, which is available in ProM version 6.4.
A ProM plug-in requires a number of input objects and produces one or more output
objects. The main input object of our plug-in is an event log, whereas the output is a
decision tree. To build decision trees, the plug-in leverages on the implementation of
the C4.5 algorithm in Weka (http://weka.sourceforge.net/). As mentioned before, our
framework envisions the possibility to augment/manipulate the event logs with
additional features. On the this concern, the tool is easily extensible: a new log
manipulation can be easily plugged in by (1) implementing 3 methods in a Java class that inherits
from an abstract class and (2) programmatically adding it to a given Java set of available
log manipulations. To date, the implementation already includes an extensive number
of manipulations, which cover different process perspectives (time, control-flow, data,
resource and conformance) and are listed in Table 1 of [
Usage of the Tool to Perform a Correlation Analysis Use Case
In [
The analysis use case aims at correlating the duration of executing activity Afspraak
(in Dutch, appointment) to other process characteristics. This activity is performed by
physicians who periodically visit hospitalized patients. After starting ProM, the user
needs to choose plug-in Perform Prediction of Business Process Features. In addition
to giving an event log as input, we also put forward a second object that provides the
necessary information to augment/manipulate events with characteristics linked to the
conformance of process instances against a prescribed process model (see [
By passing over the labels with the mouse, different configuration panels are shown (see Figure 2) The first step concerns with choosing the characteristics to consider: Figure 2(a) shows the panel where users select the characteristics to consider among those available. These characteristics are visualized in a tree and grouped by the process perspective to which they refer. By selecting a node in a tree, characteristics are added to those to consider.
The characteristics linked to conformance are displayed differently: by selecting
Consider fitness as feature, each event is augmented with the level of fitness of the trace
to which the event belongs. By clicking on Open the fitness frame, users
(a) Panel to select the process characteristics to consider.
(b) Panel to filter on the activities of the events to retain.
can selectively decide (panel
not shown here) if the number
of deviations for certain single
activities should be considered
as characteristics (see [
The filtering of events happens in the phase that follows the manipulation with additional characteristics. This means that the choice of events to retain does not influence how events are augmented with additional characteristics, e.g. referring to the number of executions of given activities or to the previous/next activity in trace. As final step, the analyst needs to choose which characteristic is the dependent one. This is done through the panel Configuration, shown in Figure 2(c). For our analysis use case, we selected Activity Duration as dependent characteristic. pruned, with the constraint that no less than 167 events can be associated with a leaf so as to balance under- and over-fitting problems. C4.5 requires a dependent characteristic to be discrete. The activity duration is a continuous characteristic and, hence, needs to be discretized before being used. Different discretization techniques are accessible through the Discretization panel (not shown here). For this analysis, we opted for equal-frequency binning: intervals are of different sizes but (roughly) the same number of observed values falls into each one.