Process Mining is a research discipline that aims to analyze business processes based on event logs. The event logs are among others used to create models for predicting the next activity of a given process instance. Existing models use Bayesian Networks or Markov Chains to predict the next activity in a work ow. These models require knowledge about the occurence of activities in the business process, which is usually based on expert knowledge or based on previous work ows from event logs. Based on previous work, we will i) represent a business process in sBPMN and extend our annotation tool to ii) compute the likelihood of occurrence of activities in a business process and check for stochastic dependency in a process and iii) use the generated knowledge to annotate the business process.
A very common research topic in Process Mining is the prediction of next activities in a business process. The recommendation of next activities can either be based on a schema, if a target process has been de ned, on statistical methods, which exploits knowledge from past events, stored in a repository, or as a hybrid system, both in combination. The persons, involved in the process, get recommendations from a system, which next activities they should perform. Often are those recommender systems out to optimize a certain optimization function, which aims at decreasing the mortality of patients, runtime of business process or increase the satisfaction of involved persons. However, it is often very hard to predict next activities, because the underlying data is very heterogeneous and di erent linguistic explanations for similar activities lead to fault predictions. In addition, if no target process is de ned, then the number of process variations might be high due to di erent opinions in the process execution by various persons. This is among others given in the medical domain. Various physicians might have di erent opinions and follow di erent best practices. This leads to a high number of process variations for similar processes. A further impact in predicting next activities is the various number of in uences that comes in.
Semantics and background knowledge might improve the results of predicting next activities in a process. So far, including semantics and background knowledge, has not been considered deeply. Often, only the sequence of activities has been considered. However, there is semantics hidden in the log les of processes that can be used to improve statistical methods to predict next activities. Often occurrence of activities depend on each other.
Our previous works focused on nding correlations in meta-information of a
business process. Now we are interested in nding dependencies in the worfklow
of a business process. Our aim is to compute the likelihood of the next activity
and therefore the likelihood of a certain outcome and work ow. We will also check
if events in a business process are stochastically independent. This knowledge
can be used to compute the likelihood of di erent work ows and predicting the
next activity. By annotating the business process with the likelihoods of their
occurrences, a Markov chain can easily be created [
Based on previous work, we will represent a business process in sBPMN and i) extend our annotation tool to ii) compute the likelihood of occurrence of activities in a business process and check for stochastic dependency in a process and iii) use the generated knowledge to annotate the business process.
In summary, this paper makes the following contributions: 1. Compute the likelihood of occurrence of activities in a business process and check for stochastically dependencies. 2. Annotate the sBPMN with the generated knowledge.
The remainder of this paper is structured as follows: the following section (Section 2) introduces the methods used to compute the likelihoods and stochastic dependencies. Section 3 demonstrates the practical applicability of our solution by realizing our approach with open-source process data and evaluating the added value to recommend next activities and computing the likelihood of certain work ows. Related work is described in section 4. We sum up our contributions and provide conclusions and future work in section 5.
Our aim is to annotate activities of a business process with their corresponding
likelihood of occurrence. This knowledge can be used to compute the
likelihood of a certain outcome, work ow and use this knowledge in predicting the
next activity. We start from the premise that a target process is already
dened in Business Process Model and Notation (BPMN). However, because we
want to exploit semantic relationship and meta-information in future, we are
interested in including the semantic information into BPMN. Semantic Business
Process Model and Notation (sBPMN) allows to add meaning in form of
metainformation and background knowledge to each process elements. The result is
a machine-readable format, which allows for reasoning on the process
description. Therefore, we will use sBPMN to model the processes, because it allows to
add semantics to business processes, which is used to describe the information
of the business processes. Hence, a rst step is to transform a given business
process in BPMN into sBPMN. For this purpose, we will use previous work that
automatically transforms a BPMN process in the standard format BPMN 2.0
XML by OMG [
Critical issues in a process are usually branches in which involved persons have to decide which way to pick. If no decision criteria is given, then a very basic approach is to assume a Laplace's probability space for the next activity. This means that the likelihood of occurrence of each previous activity is the same (uniform distribution over the following activities). We exemplary depicted in gure 2 the likelihoods of the next activity of the branches by using a Laplace's probability space assumption. For all other previous activities, without branches, is the likelihood 1.
However, the assumption of a Laplace's probability space is for recommending next activities in a real-world scenario not su cient. To improve our model, we will compute the likelihood of occurrence by using existing process instances from a repository. For this purpose, we use historic data from a repository to compute the likelihood of the next activities. For sequences of activities, the next activity is obviously. Therefore, the likelihood of the next activity is 1. So crucial parts of the process are the branches for which multiple activities can follow. In our example those are the activities after Check application form completeness, Assess eligibility, Check if home insurance quote is requested and Verify repayment agreement.
By using the computed likelihoods, one can easily use them to forecast the next activity and to model a Markov chain. The assumption of Markov chains is that the next state depends only on the current state. Which was calculated above. Therefore, we can use to predict the outcome of a work ow and the likelihood of a certain work ow by using the Markov chain assumption. However, the outcome or occurrence of activities might dependent not only on the current activities or states but on previous activities. E.g. one might think that the occurrence of returning the application back to the applicant due to an incomplete form might lead to a higher likelihood of rejecting the application. In order to check such dependencies, we will check for stochastically dependencies of activities in the process. Two events A and B are stochastically independent if the following applies:
P (A) P (B) = P (A \ B) 3
We used a free available data set from BPI Challenge 2012 [7]. The log les
of each process instance were created synthetically. We used the parallel target
process, which contains 10,000 process instances. Meta-information, except for
time-stamps, are not given. We described the target process by using
Cognitive Process Designer [8, 9]. This tool is an extension to Semantic MediaWiki
(SMW) [10] that allows for capturing BPMN diagrams, meta-information about
the activities and describe the information semantically. Semantic MediaWiki is
a powerful collaborative knowledge management system, using the MediaWiki
engine and allowing for capturing information in a structured way. The captured
information are stored by using RDF as standard format and can therefore be
queried. By using an appropriate ontology, the target process is available in
sBPMN and can easily be enriched with further information. We used an
ontology published by DKM [
To comprehend the application, we will show two examples in the following. In total 1,070 times were the Activity Return application back to applicant chosen as next activity after Check application form completeness and 10,000 times Appraise property. This leads to the following likelihoods:
P (Xt = Appraise PropertyjXt 1 = Check application) = P (Xt = Return applicationjXt 1 = Check application) =
For the next decision point (Asses eligibility ), we computed the following likelihoods: P (Xt = Reject applicationjXt 1 = Asses eligibility) = P (Xt = Prepare acceptance packjXt 1 = Asses eligibility) = = 0:5084
The bash script computed stochastically dependency between activities as well. Therefore, it checks if there might be dependencies between the occurence of certain activities exist. For instance, one might think of a stochastic dependency if an application were returned due to an incompletness of the form and a rejection. Therefore we checked the stochastically dependency between these two events.
Because 0.0473 0.0486, we can assume that the events Return application and Reject application are stochastically independent. We checked other combinations of events for stochastically dependency as well. E.g. we checked if the events Return application and Cancel application are stochastically independent. Due to the fact that 0.0236 0.0223, we assume that there is no dependency.
We performed tests by using a sampling set of 9,000 work ows for computing the likelihoods of the next activities and predicted the next activities by using the likelihoods. As assumed, because we could not nd any dependency in the workows, the likelihoods also re ect the error rate. E.g. we computed the likelihood P (Appraise PropertyjCheck application form completness) = respectively P (Return application back to applicantjCheck application form completness) = 0:0962. Therefore, as a very basic approach, we always suggested the activity with the highest likelihood for the evaluation set (1,000 work ows). In this case, the error rate was 0.1007. Which is the proportion of the likelihood of the total result ((1000 0:1007 + 9000 0:0962) 10000).
We used the generated knowledge about the likelihood of occurence of each event and the likelihood of the next event in our sBPMN model by including it as meta-information. We attached these information on the activities and decision nodes. By using the likelihoods and assuming a Markov chain characteristic, we can calculate the likelihood of di erent work ows of the process. E.g. the likelihood that the the application is returned once to the applicant and then rejected is 0:0429. This corresponds to the likelihood of this scenario given in the data (435 times occured this work ow1). The respective likelihood that the application form is returned two times to the applicant and then rejected is 0:0042, which also corresponds to the occurred likelihood in the data (47 times occured this work ow2). 4
Our approach is addressed by roughly three kinds of work: 1) Match BPMN process to sBPMN, 2) computing likelihoods for next activities based on historic data and 3) annotating business processes with the generated knowledge.
sBPMN was developed to allow for extending BPMN elements with
additional information and background knowledge to enhance analysis [11, 12]. So
far, existing work already addressed the transformation of BPMN into other
languages like e.g. BPEL [13, 14]. Further work developed ontologies to
semantically enrich BPMN in sBPMN [
Another aspect that is tackled in our approach is the computation of the
occurrence of activities in a process in general, but also conditioned on
previous activites, based on historic work ows. Finding predictive models to forecast
activities is in fact a very prominent example in Process Mining. Forecasting
tools exist that on the one hand detect data attributes that in uence the choices
in a process [15], as well as to detect decision points and try to minimize
uncertainties in a process [16]. Existing approaches forecast the next activity and
activity durations in a process by using decision trees and rule induction [17, 18],
regression [19] or a classi cation model to support prediction of activities [20]
and exceptions [
The last addressed topic is the enrichment of the business process with the
generated knowledge. Current work used semantic information to increase the
precision of process models [
In this paper we present an approach to map a BPMN process into sBPMN. We used the work ows of the process to calculate the likelihood of occurence of each activity and the likelihood of occurence for each next activity, depending on the current activity. This generated knowledge could in turn be used to enrich the sBPMN with further information. We used the likelihood of occurence of the activities to check for stochastically dependecies. In addition, the likelihoods were used to compute the likelihood of various work ows. Thereby, we assumed the Markov chain characteristic of the process. Future work includes combining our previous work of detecting correlations of meta-information and the outcome of this work. We will combine the detection of correlation of meta-information and dependency of the work ow to detect crucial parts of the process and detect unknown in uences in the process. We suppose to indicate the critera in decision points.