[Workflow Mining] Discovering Process. Models From Event Logs

Workflow Mining: Discovering process models from event logs
W.M.P. van der Aalst, A.J.M.M. Weijters, and L. Maruster
Department of Technology Management, Eindhoven University of Technology P.O. Box 513, NL-5600 MB, Eindhoven, The Netherlands. [email protected]

Abstract. Contemporary workflow management systems are driven by explicit process models, i.e., a completely specified workflow design is required in order to enact a given workflow process. Creating a workflow design is a complicated time-consuming process and typically there are discrepancies between the actual workflow processes and the processes as perceived by the management. Therefore, we have developed techniques for discovering workflow models. Starting point for such techniques is a so-called “workflow log” containing information about the workflow process as it is actually being executed. We present a new algorithm to extract a process model from such a log and represent it in terms of a Petri net. However, we will also demonstrate that it is not possible to discover arbitrary workflow processes. In this paper we explore a class of workflow processes that can be discovered. We show that the α-algorithm can successfully mine any workflow represented by a so-called SWF-net. Key words: Workflow mining, Workflow management, Data mining, Petri nets.

1

Introduction

During the last decade workflow management concepts and technology [3, 5, 15, 26, 28] have been applied in many enterprise information systems. Workflow management systems such as Staffware, IBM MQSeries, COSA, etc. offer generic modeling and enactment capabilities for structured business processes. By making graphical process definitions, i.e., models describing the life-cycle of a typical case (workflow instance) in isolation, one can configure these systems to support business processes. Besides pure workflow management systems many other software systems have adopted workflow technology. Consider for example ERP (Enterprise Resource Planning) systems such as SAP, PeopleSoft, Baan and Oracle, CRM (Customer Relationship Management) software, etc. Despite its promise, many problems are encountered when applying workflow technology. One of the problems is that these systems require a workflow design, i.e., a designer has to construct a detailed model accurately describing the routing of work. Modeling a workflow is far from trivial: It requires deep knowledge of the workflow language and lengthy discussions with the workers and management involved.

Instead of starting with a workflow design, we start by gathering information about the workflow processes as they take place. We assume that it is possible to record events such that (i) each event refers to a task (i.e., a well-defined step in the workflow), (ii) each event refers to a case (i.e., a workflow instance), and (iii) events are totally ordered (i.e., in the log events are recorded sequentially even though tasks may be executed in parallel). Any information system using transactional systems such as ERP, CRM, or workflow management systems will offer this information in some form. Note that we do not assume the presence of a workflow management system. The only assumption we make, is that it is possible to collect workflow logs with event data. These workflow logs are used to construct a process specification which adequately models the behavior registered. We use the term process mining for the method of distilling a structured process description from a set of real executions.
case identifier task identifier case 1 task A case 2 task A case 3 task A case 3 task B case 1 task B case 1 task C case 2 task C case 4 task A case 2 task B case 2 task D case 5 task A case 4 task C case 1 task D case 3 task C case 3 task D case 4 task B case 5 task E case 5 task D case 4 task D Table 1. A workflow log.

To illustrate the principle of process mining, we consider the workflow log shown in Table 1. This log contains information about five cases (i.e., workflow instances). The log shows that for four cases (1,2,3, and 4) the tasks A, B, C, and D have been executed. For the fifth case only three tasks are executed: tasks A, E, and D. Each case starts with the execution of A and ends with the execution of D. If task B is executed, then also task C is executed. However, for some cases task C is executed before task B. Based on the information shown in Table 1 and by making some assumptions about the completeness of the log (i.e., assuming that the cases are representative and a sufficient large subset 2

of possible behaviors is observed), we can deduce for example the process model shown in Figure 1. The model is represented in terms of a Petri net [39]. The Petri net starts with task A and finishes with task D. These tasks are represented by transitions. After executing A there is a choice between either executing B and C in parallel or just executing task E. To execute B and C in parallel two non-observable tasks (AND-split and AND-join) have been added. These tasks have been added for routing purposes only and are not present in the workflow log. Note that we assume that two tasks are in parallel if they appear in any order. However, by distinguishing between start events and end events for tasks it is possible to explicitly detect parallelism. Start events and end events can also be used to indicate that tasks take time. However, to simplify the presentation we assume tasks to be atomic without losing generality. In fact in our tool EMiT [4] we refine this even further and assume a customizable transaction model for tasks involving events like “start task”, “withdraw task”, “resume task”, “complete task”, etc. [4]. Nevertheless, it is important to realize that such an approach only works if events these are recorded at the time of their occurrence.

B
AND -split AND -join

C A E D

Fig. 1. A process model corresponding to the workflow log.

The basic idea behind process mining, also referred to as workflow mining, is to construct Figure 1 from the information given in Table 1. In this paper, we will present a new algorithm and prove its correctness. Process mining is useful for at least two reasons. First of all, it could be used as a tool to find out how people and/or procedures really work. Consider for example processes supported by an ERP system like SAP (e.g., a procurement process). Such a system logs all transactions but in many cases does not enforce a specific way of working. In such an environment, process mining could be used to gain insight in the actual process. Another example would be the flow of patients in a hospital. Note that in such an 3

environment all activities are logged but information about the underlying process is typically missing. In this context it is important to stress that management information systems provide information about key performance indicators like resource utilization, flow times, and service levels but not about the underlying business processes (e.g., causal relations, ordering of activities, etc.). Second, process mining could be used for Delta analysis, i.e., comparing the actual process with some predefined process. Note that in many situations there is a descriptive or prescriptive process model. Such a model specifies how people and organizations are assumed/expected to work. By comparing the descriptive or prescriptive process model with the discovered model, discrepancies between both can be detected and used to improve the process. Consider for example the so-called reference models in the context of SAP. These models describe how the system should be used. Using process mining it is possible to verify whether this is the case. In fact, process mining could also be used to compare different departments/organizations using the same ERP system. An additional benefit of process mining is that information about the way people and/or procedures really work and differences between actual processes and predefined processes can be used to trigger Business Process Reengineering (BPR) efforts or to configure “process-aware information systems” (e.g., workflow, ERP, and CRM systems). Table 1 contains the minimal information we assume to be present. In many applications, the workflow log contains a timestamp for each event and this information can be used to extract additional causality information. Moreover, we are also interested in the relation between attributes of the case and the actual route taken by a particular case. For example, when handling traffic violations: Is the make of a car relevant for the routing of the corresponding traffic violations? (E.g., People driving a Ferrari always pay their fines in time.) For this simple example, it is quite easy to construct a process model that is able to regenerate the workflow log. For larger workflow models this is much more difficult. For example, if the model exhibits alternative and parallel routing, then the workflow log will typically not contain all possible combinations. Consider 10 tasks which can be executed in parallel. The total number of interleavings is 10! = 3628800. It is not realistic that each interleaving is present in the log. Moreover, certain paths through 4

the process model may have a low probability and therefore remain undetected. Noisy data (i.e., logs containing rare events, exceptions and/or incorrectly recorded data) can further complicate matters. In this paper, we do not focus on issues such as noise. We assume that there is no noise and that the workflow log contains “sufficient” information. Under these ideal circumstances we investigate whether it is possible to rediscover the workflow process, i.e., for which class of workflow models is it possible to accurately construct the model by merely looking at their logs. This is not as simple as it seems. Consider for example the process model shown in Figure 1. The corresponding workflow log shown in Table 1 does not show any information about the AND-split and the AND-join. Nevertheless, they are needed to accurately describe the process. These and other problems are addressed in this paper. For this purpose we use workflow nets (WF-nets). WF-nets are a class of Petri nets specifically tailored towards workflow processes. Figure 1 shows an example of a WF-net.

WF-net workflow log

WF-net

WF1
generate workflow log based on WF-net construct WF-net based on applying workflow mining techniques

WF2

WF1 = WF2 ?

Fig. 2. The rediscovery problem: For which class of WF-nets is it guaranteed that WF 2 is equivalent to WF 1 ?

To illustrate the rediscovery problem we use Figure 2. Suppose we have a log based on many executions of the process described by a WF-net WF 1 . Based on this workflow log and using a mining algorithm we construct a WF-net WF 2 . An interesting question is whether WF 1 = WF 2 . In this paper, we explore the class of WF-nets for which WF 1 = WF 2 . Note that the rediscovery problem is only addressed to explore the theoretical limits of process mining and to test the algorithm presented in this paper. We have used 5

these results to develop tools that can discover unknown processes and have successfully applied these tools to mine real processes. The remainder of this paper is organized as follows. First, we introduce some preliminaries, i.e., Petri nets and WF-nets. In Section 3 we formalize the problem addressed in this paper. Section 4 discusses the relation between causality detected in the log and places connecting transitions in the WF-net. Based on these results, an algorithm for process mining is presented. The quality of this algorithm is supported by the fact that it is able to rediscover a large class of workflow processes. The paper finishes with an overview of related work and some conclusions.

2

Preliminaries

This section introduces the techniques used in the remainder of this paper. First, we introduce standard Petri-net notations, then we define the class of WF-nets. 2.1 Petri nets

We use a variant of the classic Petri-net model, namely Place/Transition nets. For an elaborate introduction to Petri nets, the reader is referred to [12, 37, 39]. Definition 2.1. (P/T-nets)1 An Place/Transition net, or simply P/T-net, is a tuple (P, T, F ) where: 1. P is a finite set of places, 2. T is a finite set of transitions such that P ∩ T = ∅, and 3. F ⊆ (P × T ) ∪ (T × P ) is a set of directed arcs, called the flow relation. A marked P/T-net is a pair (N, s), where N = (P, T, F ) is a P/T-net and where s is a bag over P denoting the marking of the net. The set of all marked P/T-nets is denoted N. A marking is a bag over the set of places P , i.e., it is a function from P to the natural numbers. We use square brackets for the enumeration of a bag, e.g., [a2 , b, c3 ] denotes the bag with two a-s, one b, and three c-s. The sum of two bags (X + Y ), the difference
1

In the literature, the class of Petri nets introduced in Definition 2.1 is sometimes referred to as the class of (unlabeled) ordinary P/T-nets to distinguish it from the class of Petri nets that allows more than one arc between a place and a transition.

6

(X − Y ), the presence of an element in a bag (a ∈ X), and the notion of subbags (X ≤ Y ) are defined in a straightforward way and they can handle a mixture of sets and bags. Let N = (P, T, F ) be a P/T-net. Elements of P ∪ T are called nodes. A node x is an input node of another node y iff there is a directed arc from x to y (i.e., (x, y) ∈ F ). Node x is an output node of y iff (y, x) ∈ F . For any x ∈ P ∪ T ,
N

•

x = {y | (x, y) ∈ F }

and x N = {y | (x, y) ∈ F }; the superscript N may be omitted if clear from the context. • Figure 1 shows a P/T-net consisting of 8 places and 7 transitions. Transition A has one input place and one output place, transition AND-split has one input place and two output places, and transition AND-join has two input places and one output place. The black dot in the input place of A represents a token. This token denotes the initial marking. The dynamic behavior of such a marked P/T-net is defined by a firing rule. Definition 2.2. (Firing rule) Let (N = (P, T, F ), s) be a marked P/T-net. Transition t ∈ T is enabled, denoted (N, s)[t , iff •t ≤ s. The firing rule (N, s)[t ⇒ (N, s) [t (N, s − •t + t•). In the marking shown in Figure 1 (i.e., one token in the source place), transition A is enabled and firing this transition removes the token from the input place and puts a token in the output place. In the resulting marking, two transitions are enabled: E and AND-split. Although both are enabled only one can fire. If AND-split fires, one token is consumed and two tokens are produced. Definition 2.3. (Reachable markings) Let (N, s0 ) be a marked P/T-net in N . A marking s is reachable from the initial marking s0 iff there exists a sequence of enabled transitions whose firing leads from s0 to s. The set of reachable markings of (N, s0 ) is denoted [N, s0 . The marked P/T-net shown in Figure 1 has 8 reachable markings. Sometimes it is convenient to know the sequence of transitions that are fired in order to reach some given marking. This paper uses the following notations for sequences. Let A be some alphabet of identifiers. A sequence of length n, for some natural number n ∈ IN, over alphabet A is a function σ : {0, . . . , n − 1} → A. The sequence of length zero is called 7 [ ⊆ N ×T ×N is the smallest relation satisfying for any (N = (P, T, F ), s) ∈ N and any t ∈ T ,

the empty sequence and written ε. For the sake of readability, a sequence of positive length is usually written by juxtaposing the function values: For example, a sequence σ = {(0, a), (1, a), (2, b)}, for a, b ∈ A, is written aab. The set of all sequences of arbitrary length over alphabet A is written A∗ . Definition 2.4. (Firing sequence) Let (N, s0 ) with N = (P, T, F ) be a marked P/T net. A sequence σ ∈ T ∗ is called a firing sequence of (N, s0 ) iff, for some natural number n ∈ IN, there exist markings s1 , . . . , sn and transitions t1 , . . . , tn ∈ T such that σ = t1 . . . tn and, for all i with 0 ≤ i < n, (N, si )[ti+1 and si+1 = si − •ti+1 + ti+1•. (Note that n = 0 implies that σ = ε and that ε is a firing sequence of (N, s0 ).) Sequence σ is said to be enabled in marking s0 , denoted (N, s0 )[σ . Firing the sequence σ results in a marking sn , denoted (N, s0 ) [σ (N, sn ). Definition 2.5. (Connectedness) A net N = (P, T, F ) is weakly connected, or simply connected, iff, for every two nodes x and y in P ∪ T , x(F ∪ F −1 )∗ y, where R−1 is the inverse and R∗ the reflexive and transitive closure of a relation R. Net N is strongly connected iff, for every two nodes x and y, xF ∗ y. We assume that all nets are weakly connected and have at least two nodes. The P/T-net shown in Figure 1 is connected but not strongly connected because there is no directed path from the sink place to the source place, or from D to A, etc. Definition 2.6. (Boundedness, safeness) A marked net (N = (P, T, F ), s) is bounded iff the set of reachable markings [N, s is finite. It is safe iff, for any s ∈ [N, s and any p ∈ P , s (p) ≤ 1. Note that safeness implies boundedness. The marked P/T-net shown in Figure 1 is safe (and therefore also bounded) because none of the 8 reachable states puts more than one token in a place. Definition 2.7. (Dead transitions, liveness) Let (N = (P, T, F ), s) be a marked P/T-net. A transition t ∈ T is dead in (N, s) iff there is no reachable marking s ∈ [N, s such that (N, s )[t . (N, s) is live iff, for every reachable marking s ∈ [N, s and t ∈ T , there is a reachable marking s ∈ [N, s such that (N, s )[t . Note that liveness implies the absence of dead transitions. None of the transitions in the marked P/T-net shown in Figure 1 is dead. However, the marked P/T-net is not live since it is not possible to enable each transition continuously. 8

2.2

Workflow nets

Most workflow systems offer standard building blocks such as the AND-split, AND-join, OR-split, and OR-join [5, 15, 26, 28]. These are used to model sequential, conditional, parallel and iterative routing (WFMC [15]). Clearly, a Petri net can be used to specify the routing of cases. Tasks are modeled by transitions and causal dependencies are modeled by places and arcs. In fact, a place corresponds to a condition which can be used as preand/or post-condition for tasks. An AND-split corresponds to a transition with two or more output places, and an AND-join corresponds to a transition with two or more input places. OR-splits/OR-joins correspond to places with multiple outgoing/ingoing arcs. Given the close relation between tasks and transitions we use the terms interchangeably. A Petri net which models the control-flow dimension of a workflow, is called a WorkFlow net (WF-net). It should be noted that a WF-net specifies the dynamic behavior of a single case in isolation. ¯ Definition 2.8. (Workflow nets) Let N = (P, T, F ) be a P/T-net and t a fresh identifier not in P ∪ T . N is a workflow net (WF-net) iff: 1. object creation: P contains an input place i such that •i = ∅, 2. object completion: P contains an output place o such that o• = ∅, ¯ ¯ ¯ ¯ 3. connectedness: N = (P, T ∪ {t}, F ∪ {(o, t), (t, i)}) is strongly connected, The P/T-net shown in Figure 1 is a WF-net. Note that although the net is not strongly ¯ ¯ ¯ ¯ connected, the short-circuited net N = (P, T ∪ {t}, F ∪ {(o, t), (t, i)}) (i.e., the net ¯ with transition t connecting o to i) is strongly connected. Even if a net meets all the syntactical requirements stated in Definition 2.8, the corresponding process may exhibit errors such as deadlocks, tasks which can never become active, livelocks, garbage being left in the process after termination, etc. Therefore, we define the following correctness criterion. Definition 2.9. (Sound) Let N = (P, T, F ) be a WF-net with input place i and output place o. N is sound iff: 1. safeness: (N, [i]) is safe, 2. proper completion: for any marking s ∈ [N, [i] , o ∈ s implies s = [o], 3. option to complete: for any marking s ∈ [N, [i] , [o] ∈ [N, s , and 9

4. absence of dead tasks: (N, [i]) contains no dead transitions. The set of all sound WF-nets is denoted W. The WF-net shown in Figure 1 is sound. Soundness can be verified using standard Petrinet-based analysis techniques. In fact soundness corresponds to liveness and safeness of the corresponding short-circuited net [1, 2, 5]. This way efficient algorithms and tools can be applied. An example of a tool tailored towards the analysis of WF-nets is Woflan [47].

3

The rediscovery problem

After introducing some preliminaries we return to the topic of this paper: workflow mining. The goal of workflow mining is to find a workflow model (e.g., a WF-net) on the basis of a workflow log. Table 1 shows an example of a workflow log. Note that the ordering of events within a case is relevant while the ordering of events amongst cases is of no importance. Therefore, we define a workflow log as follows. Definition 3.1. (Workflow trace, Workflow log) Let T be a set of tasks. σ ∈ T ∗ is a workflow trace and W ∈ P(T ∗ ) is a workflow log.2 The workflow trace of case 1 in Table 1 is ABCD. The workflow log corresponding to Table 1 is {ABCD, ACBD, AED}. Note that in this paper we abstract from the identity of cases. Clearly the identity and the attributes of a case are relevant for workflow mining. However, for the theoretical results in this paper, we can abstract from this. For similar reasons, we abstract from the frequency of workflow traces. In Table 1 workflow trace ABCD appears twice (case 1 and case 3), workflow trace ACBD also appears twice (case 2 and case 4), and workflow trace AED (case 5) appears only once. These frequencies are not registered in the workflow log {ABCD, ACBD, AED}. Note that when dealing with noise, frequencies are of the utmost importance. However, in this paper we do not deal with issues such as noise. Therefore, this abstraction is made to simplify notation. For readers interested in how we deal with noise and related issues, we refer to [31, 32, 48–50]
2

P(T ∗ ) is the powerset of T ∗ , i.e., W ⊆ T ∗ .

10

To find a workflow model on the basis of a workflow log, the log should be analyzed for causal dependencies, e.g., if a task is always followed by another task it is likely that there is a causal relation between both tasks. To analyze these relations we introduce the following notations. Definition 3.2. (Log-based ordering relations) Let W be a workflow log over T , i.e., W ∈ P(T ∗ ). Let a, b ∈ T : – a >W b iff there is a trace σ = t1 t2 t3 . . . tn−1 and i ∈ {1, . . . , n − 2} such that σ ∈ W and ti = a and ti+1 = b, – a →W b iff a >W b and b >W a, – a#W b iff a >W b and b >W a, and – a
Wb

iff a >W b and b >W a.

Consider the workflow log W = {ABCD, ACBD, AED} (i.e., the log shown in Table 1). Relation >W describes which tasks appeared in sequence (one directly following the other). Clearly, A >W B, A >W C, A >W E, B >W C, B >W D, C >W B, C >W D, and E >W D. Relation →W can be computed from >W and is referred to as the (direct) causal relation derived from workflow log W . A →W B, A →W C, A →W E, B →W D, C →W D, and E →W D. Note that B →W C because C >W B. Relation and C
W B. W

suggests
WC

potential parallelism. For log W tasks B and C seem to be in parallel, i.e., B

If two tasks can follow each other directly in any order, then all possible

interleavings are present and therefore they are likely to be in parallel. Relation #W gives pairs of transitions that never follow each other directly. This means that there are no direct causal relations and parallelism is unlikely. Property 3.3. Let W be a workflow log over T . For any a, b ∈ T : a →W b or b →W a or a#W b or a
W b.

Moreover, the relations →W , →−1 , #W , and W

W

are mutually exclusive

and partition T × T .3 This property can easy be verified. Note that →W = (>W \ >−1 ), →−1 = (>−1 \ >W ), W W W #W = (T × T ) \ (>W ∪ >−1 ), W #W ∪
3

W

= (>W ∩ >−1 ). Therefore, T × T = →W ∪ →−1 ∪ W W

W.

If no confusion is possible, the subscript W is omitted.

To simplify the use of logs and sequences we introduce some additional notations.
→−1 is the inverse of relation →W , i.e., →−1 = {(y, x) ∈ T × T | x →W y}. W W

11

Definition 3.4. (∈, first, last) Let A be a set, a ∈ A, and σ = a1 a2 . . . an ∈ A∗ a sequence over A of length n. ∈, first, last are defined as follows: 1. a ∈ σ iff a ∈ {a1 , a2 , . . . an }, 2. first(σ) = a1 , if n ≥ 1, and 3. last(σ) = an , if n ≥ 1. To reason about the quality of a workflow mining algorithm we need to make assumptions about the completeness of a log. For a complex process, a handful of traces will not suffice to discover the exact behavior of the process. Relations →W , →−1 , #W , and W
W

will be crucial information for any workflow-mining algorithm. Since these relations

can be derived from >W , we assume the log to be complete with respect to this relation. Definition 3.5. (Complete workflow log) Let N = (P, T, F ) be a sound WF-net, i.e., N ∈ W. W is a workflow log of N iff W ∈ P(T ∗ ) and every trace σ ∈ W is a firing sequence of N starting in state [i] and ending in [o], i.e., (N, [i])[σ (N, [o]). W is a complete workflow log of N iff (1) for any workflow log W of N : >W ⊆>W , and (2) for any t ∈ T there is a σ ∈ W such that t ∈ σ. A workflow log of a sound WF-net only contains behaviors that can be exhibited by the corresponding process. A workflow log is complete if all tasks that potentially directly follow each other in fact directly follow each other in some trace in the log. Note that transitions that connect the input place i of a WF-net to its output place o are “invisible” for >W . Therefore, the second requirement has been added. If there are no such transitions, this requirement can be dropped as is illustrated by the following property. Property 3.6. Let N = (P, T, F ) be a sound WF-net. If W is a complete workflow log of N , then {t ∈ T | ∃t ∈T t >W t ∨ t >W t} = {t ∈ T | t ∈ i • ∩ • o}. Proof. Consider a transition t ∈ T . Since N is sound there is firing sequence containing t. If t ∈ i • ∩ • o, then this sequence has length 1 and t cannot appear in >W because this is the only firing sequence containing t. If t ∈ i • ∩ • o, then the sequence has at least length 2, i.e., t is directly preceded or followed by a transition and therefore appears in >W . ✷

The definition of completeness given in Definition 3.5 may seem arbitrary but it is not. Note that it would be unrealistic to assume that all possible firing sequences are present 12

in the log. First of all, the number of possible sequences may be infinite (in case of loops). Second, parallel processes typically have an exponential number of states and, therefore, the number of possible firing sequences may be enormous. Finally, even if there is no parallelism and no loops but just N binary choices, the number of possible sequences may be 2N . Therefore, we need a weaker notion of completeness. If there is no parallelism and no loops but just N binary choices, the number of cases required may be as little as 2. Of course for large N it is unlikely that all choices are observed in just 2 cases but still it indicates that this requirement is considerably less demanding than observing all possible sequences. The same holds for processes with loops and parallelism. If a process has N sequential fragments which each exhibit parallelism, the number of cases needed to observe all possible combinations is exponential in the number of fragments. Using our notion of completeness, this is not the case. One could consider even weaker notions of completeness, however, as will be shown in the remainder, even this notion of completeness (i.e., Definition 3.5) is in some situations too weak to detect certain advanced routing patterns. We will formulate the rediscovery problem introduced in Section 1 assuming a complete workflow log as described in Definition 3.5. Before formulating this problem we define what it means for a WF-net to be rediscovered. Definition 3.7. (Ability to rediscover) Let N = (P, T, F ) be a sound WF-net, i.e., N ∈ W, and let α be a mining algorithm which maps workflow logs of N onto sound WF-nets, i.e., α : P(T ∗ ) → W. If for any complete workflow log W of N the mining algorithm returns N (modulo renaming of places), then α is able to rediscover N . Note that no mining algorithm is able to find names of places. Therefore, we ignore place names, i.e., α is able to rediscover N iff α(W ) = N modulo renaming of places. The goal of this paper is twofold. First of all, we are looking for a mining algorithm that is able to rediscover sound WF-nets, i.e., based on a complete workflow log the corresponding workflow process model can be derived. Second, given such an algorithm we want to indicate the class of workflow nets which can be rediscovered. Clearly, this class should be as large as possible. Note that there is no mining algorithm which is able to rediscover all sound WF-nets. For example, if in Figure 1 we add a place p connecting transitions A and D, there is no mining algorithm able to detect p since this place is 13

implicit, i.e., the addition of the place does not change the behavior of the net and therefore is not visible in the log. To conclude we summarize the rediscovery problem: “Find a mining algorithm able to rediscover a large class of sound WF-nets on the basis of complete workflow logs.” This problem was illustrated in the introduction using Figure 2.

4

Workflow mining

In this section, the rediscovery problem is tackled. Before we present a mining algorithm able to rediscover a large class of sound WF-nets, we investigate the relation between the causal relations detected in the log (i.e., →W ) and the presence of places connecting transitions. First, we show that causal relations in →W imply the presence of places. Then, we explore the class of nets for which the reverse also holds. Based on these observations, we present a mining algorithm. 4.1 Causal relations imply connecting places

If there is a causal relation between two transitions according to the workflow log, then there has to be a place connecting these two transitions. Theorem 4.1. Let N = (P, T, F ) be a sound WF-net and let W be a complete workflow log of N . For any a, b ∈ T : a →W b implies a • ∩ • b = ∅. Proof. Assume a →W b and a• ∩ •b = ∅. We will show that this leads to a contradiction and thus prove the theorem. Since a > b there is a firing sequence σ = t1 t2 t3 . . . tn−1 and i ∈ {1, . . . , n − 2} such that σ ∈ W and ti = a and ti+1 = b. Let s be the state just before firing a, i.e., (N, [i]) [σ (N, s) with σ = t1 . . . ti−1 . Let s be the marking after firing b in state s, i.e., (N, s) [b (N, s ). Note that b is enabled in s because it is enabled after firing a and a • ∩ • b = ∅ (i.e., a does not produce tokens for any of the input places of b). a cannot be enabled in s , otherwise b > a and not a →W b. Since a is enabled in s but not in s , b consumes a token from an input place of a and does not return it, i.e., ((•b) \ (b•)) ∩ •a = ∅. There is a place p such that p ∈ •a, p ∈ •b, and p ∈ b•. Moreover, a • ∩ • b = ∅. Therefore, p ∈ a•. Since the net is safe, p contains precisely one token in marking s. This token is consumed by ti = a and not returned. 14

Hence b cannot be enabled after firing ti . Therefore, σ cannot be a firing sequence of N starting in i. ✷

Let N1 = ({i, p1 , p2 , p3 , p4 , o}, {A, B, C, D}, {(i, A), (A, p1 ), (A, p2 ), (p1 , B), (B, p3 ), (p2 , C), (C, p4 ), (p3 , D), (p4 , D), (D, o)}). (This is the WF-net with B and C in parallel, see N1 in Figure 4.) W1 = {ABCD, ACBD} is a complete log over N1 . Since A →W1 B, there has to be a place between A and B. This place corresponds to p1 in N1 . Let N2 = ({i, p1 , p2 , o}, {A, B, C, D}, {(i, A), (A, p1 ), (p1 , B), (B, p2 ), (p1 , C), (C, p2 ), (p2 , D), (D, o)}). (This is the WF-net with a choice between B and C, see N2 in Figure 4.) W2 = {ABD, ACD} is a complete log over N2 . Since A →W2 B, there has to be a place between A and B. Similarly, A →W2 C and therefore there has to be a place between A and C. Both places correspond to p1 in N1 . Note that in the first example (N1 /W1 ) the two causal relations A →W1 B and A →W1 C correspond to two different places while in the second example the two causal relations A →W1 B and A →W1 C correspond to a single place. 4.2 Connecting places “often” imply causal relations

In this subsection we investigate which places can be detected by simply inspecting the log. Clearly, not all places can be detected. For example places may be implicit which means that they do not affect the behavior of the process. These places remain undetected. Therefore, we limit our investigation to WF-nets without implicit places. Definition 4.2. (Implicit place) Let N = (P, T, F ) be a P/T-net with initial marking s. A place p ∈ P is called implicit in (N, s) iff, for all reachable markings s ∈ [N, s and transitions t ∈ p•, s ≥ •t \ {p} ⇒ s ≥ •t. Figure 1 contains no implicit places. However, as indicated before, adding a place p connecting transition A and D yields an implicit place. No mining algorithm is able to detect p since the addition of the place does not change the behavior of the net and therefore is not visible in the log. For the rediscovery problem it is very important that the structure of the WF-net clearly reflects its behavior. Therefore, we also rule out the constructs shown in Figure 3. The left construct illustrates the constraint that choice and synchronization should never meet. If two transitions share an input place, and therefore “fight” for the same token, 15

Fig. 3. Two constructs not allowed in SWF-nets.

they should not require synchronization. This means that choices (places with multiple output transitions) should not be mixed with synchronizations. The right-hand construct in Figure 3 illustrates the constraint that if there is a synchronization, all preceding transitions should have fired, i.e., it is not allowed to have synchronizations directly preceded by an OR-join. WF-nets which satisfy these requirements are named structured workflow nets. Definition 4.3. (SWF-net) A WF-net N = (P, T, F ) is an SWF-net (Structured workflow net) iff: 1. For all p ∈ P and t ∈ T with (p, t) ∈ F : |p • | > 1 implies | • t| = 1. 2. For all p ∈ P and t ∈ T with (p, t) ∈ F : | • t| > 1 implies | • p| = 1. 3. There are no implicit places. At first sight the three requirements in Definition 4.3 seem quite restrictive. From a practical point of view this is not the case. First of all, SWF-nets allow for all routing constructs encountered in practice, i.e., sequential, parallel, conditional and iterative routing are possible and the basic workflow building blocks (AND-split, AND-join, ORsplit and OR-join) are supported. Second, WF-nets that are not SWF-nets are typically difficult to understand and should be avoided if possible. Third, many workflow management systems only allow for workflow processes that correspond to SWF-nets. The latter observation can be explained by the fact that most workflow management systems use a language with separate building blocks for OR-splits and AND-joins. Finally, there is a very pragmatic argument. If we drop any of the requirements stated in Definition 4.3, relation >W does not contain enough information to successfully mine all processes in the resulting class. The reader familiar with Petri nets will observe that SWF-nets belong to the class of free-choice nets [12]. This allows us to use efficient analysis techniques and advanced 16

theoretical results. For example, using these results it is possible to decide soundness in polynomial time [2]. SWF-nets also satisfy another interesting property. Property 4.4. Let N = (P, T, F ) be an SWF-net. For any a, b ∈ T and p1 , p2 ∈ P : if p1 ∈ a • ∩ • b and p2 ∈ a • ∩ • b, then p1 = p2 . This property follows directly from the definition of SWF-nets and states that no two transitions are connected by multiple places. This property illustrates that the structure of an SWF-net clearly reflects its behavior and vice versa. This is exactly what we need to be able to rediscover a WF-net from its log. We already showed that causal relations in →W imply the presence of places. Now we try to prove the reverse for the class of SWF-nets. First, we focus on the relation between the presence of places and >W . Theorem 4.5. Let N = (P, T, F ) be a sound SWF-net and let W be a complete workflow log of N . For any a, b ∈ T : a • ∩ • b = ∅ implies a >W b. Proof. See [6]. ✷

Unfortunately a• ∩ • b = ∅ does not imply a →W b. To illustrate this consider Figure 4. For the first two nets (i.e., N1 and N2 ), two tasks are connected iff there is a causal relation. This does not hold for N3 and N4 . In N3 , A →W3 B, A →W3 D, and B →W3 D. However, not B →W3 B. Nevertheless, there is a place connecting B to B. In N4 , although there are places connecting B to C and vice versa, B →W3 C and B →W3 C. These examples indicate that loops of length one (see N3 ) and length two (see N4 ) are harmful. Fortunately, loops of length three or longer are no problem as is illustrated in the following theorem. Theorem 4.6. Let N = (P, T, F ) be a sound SWF-net and let W be a complete workflow log of N . For any a, b ∈ T : a • ∩ • b = ∅ and b • ∩ • a = ∅ implies a →W b. Proof. See [6]. ✷

Acyclic nets have no loops of length one or length two. Therefore, it is easy to derive the following property. 17

B

N1

A C

D

B

N2

A C

D

B

N3

A

D

B

N4

A C

D

N5

A

B

D

E

C

Fig. 4. Five sound SWF-nets.

Property 4.7. Let N = (P, T, F ) be an acyclic sound SWF-net and let W be a complete workflow log of N . For any a, b ∈ T : a • ∩ • b = ∅ iff a →W b. The results presented thus far focus on the correspondence between connecting places and causal relations. However, causality (→W ) is just one of the four log-based ordering relations defined in Definition 4.3. The following theorem explores the relation between the sharing of input and output places and #W . Theorem 4.8. Let N = (P, T, F ) be a sound SWF-net such that for any a, b ∈ T : a • ∩ • b = ∅ or b • ∩ • a = ∅ and let W be a complete workflow log of N . 1. If a, b ∈ T and a • ∩ b• = ∅, then a#W b. 2. If a, b ∈ T and •a ∩ •b = ∅, then a#W b. 3. If a, b, t ∈ T , a →W t, b →W t, and a#W b, then a • ∩ b • ∩ • t = ∅. 4. If a, b, t ∈ T , t →W a, t →W b, and a#W b, then •a ∩ •b ∩ t• = ∅. 18

Proof. See [6]. The relations →W , →−1 , #W , and W
W

✷ are mutually exclusive. Therefore, we can derive
Wb

that for sound SWF-nets with no short loops, a t →W b, and •a ∩ •b ∩ t• = ∅, also implies a
W b.

implies a • ∩ b• = •a ∩ •b = ∅.
W b.

Moreover, a →W t, b →W t, and a • ∩ b • ∩ • t = ∅ implies a

Similarly, t →W a,

These results will be used to underpin

the mining algorithm presented in the following subsection. 4.3 Mining algorithm

Based on the results in the previous subsections we now present an algorithm for mining processes. The algorithm uses the fact that for many WF-nets two tasks are connected iff their causality can be detected by inspecting the log. Definition 4.9. (Mining algorithm α) Let W be a workflow log over T . α(W ) is defined as follows. 1. TW = {t ∈ T | ∃σ∈W t ∈ σ}, 2. TI = {t ∈ T | ∃σ∈W t = first(σ)}, 3. TO = {t ∈ T | ∃σ∈W t = last(σ)}, 4. XW = {(A, B) | A ⊆ TW ∧ B ⊆ TW ∧ ∀a∈A ∀b∈B a →W b ∧ ∀a1 ,a2 ∈A a1 #W a2 ∧ ∀b1 ,b2 ∈B b1 #W b2 }, 5. YW = {(A, B) ∈ XW | ∀(A ,B )∈XW A ⊆ A ∧ B ⊆ B =⇒ (A, B) = (A , B )}, 6. PW = {p(A,B) | (A, B) ∈ YW } ∪ {iW , oW }, 7. FW = {(a, p(A,B) ) | (A, B) ∈ YW ∧ a ∈ A} ∪ {(p(A,B) , b) | (A, B) ∈ YW ∧ b ∈ B} ∪ {(iW , t) | t ∈ TI } ∪ {(t, oW ) | t ∈ TO }, and 8. α(W ) = (PW , TW , FW ). The mining algorithm constructs a net (PW , TW , FW ). Clearly, the set of transitions TW can be derived by inspecting the log. In fact, as shown in Property 3.6, if there are no traces of length one, TW can be derived from >W . Since it is possible to find all initial transitions TI and all final transition TO , it is easy to construct the connections between these transitions and iW and oW . Besides the source place iW and the sink place oW , places of the form p(A,B) are added. For such place, the subscript refers to the set of 19

input and output transitions, i.e., •p(A,B) = A and p(A,B)• = B. A place is added inbetween a and b iff a →W b. However, some of these places need to be merged in case of OR-splits/joins rather than AND-splits/joins. For this purpose the relations XW and YW are constructed. (A, B) ∈ XW if there is a causal relation from each member of A to each member of B and the members of A and B never occur next to one another. Note that if a →W b, b →W a, or a
W b,

then a and b cannot be both in A (or B). Relation

YW is derived from XW by taking only the largest elements with respect to set inclusion. (See the end of this section for an example.)

(N3)
A

B

D

(N4)
A

B D C

Fig. 5. The α algorithm is unable to rediscover N3 and N4 .

Based on α defined in Definition 4.9, we turn to the rediscovery problem. Is it possible to rediscover WF-nets using α(W )? Consider the five SWF-nets shown in Figure 4. If α is applied to a complete workflow log of N1 , the resulting net is N1 modulo renaming of places. Similarly, if α is applied to a complete workflow log of N2 , the resulting net is N2 modulo renaming of places. As expected, α is not able to rediscover N3 and N4 (cf. Figure 5). α(W3 ) is like N3 but without the arcs connecting B to the place in-between A and D and two new places. α(W4 ) is like N4 but the input and output arc of C are removed. α(W3 ) is not a WF-net since B is not connected to the rest of the net. α(W4 ) is not a WF-net since C is not connected to the rest of the net. In both cases two arcs are missing in the resulting net. N3 and N4 illustrate that the mining algorithm is unable to deal with short loops. Loops of length three or longer are no problem. For example α(W5 ) = N5 modulo renaming of places. The following theorem proves that α is able to rediscover the class of SWF-nets provided that there are no short loops. 20

Theorem 4.10. Let N = (P, T, F ) be a sound SWF-net and let W be a complete workflow log of N . If for all a, b ∈ T a • ∩ • b = ∅ or b • ∩ • a = ∅, then α(W ) = N modulo renaming of places. Proof. Let α(W ) = (PW , TW , FW ). Since W is complete, it is easy to see that T = TW . Remains to prove that every place in N corresponds to a place in α(W ) and vice versa. Let p ∈ P . We need to prove that there is a pW ∈ PW such that
• p N = pW
NW N

•

p=

NW

•

pW and

•

. If p = i, i.e., the source place or p = o, i.e., the sink place, then it is easy

to see that there is a corresponding place in α(W ). Transitions in i N ∪ N o can fire only • • once directly at the beginning of a sequence or at the end. Therefore, the construction given in Definition 4.9 involving iW , oW , TI , and TO yields a source and sink place with identical input/output transitions. If p ∈ {i, o}, then let A =N p, B = p N , and • • pW = p(A,B) . If pW is indeed a place of α(W ), then
N

•

p=

α(W )

•

pW and p N = pW •

α(W )

•

. This

follows directly from the definition of the flow relation FW in Definition 4.9. To prove that pW = p(A,B) is a place of α(W ), we need to show that (A, B) ∈ YW . (A, B) ∈ XW , because (1) Theorem 4.6 implies that ∀a∈A ∀b∈B a →W b, (2) Theorem 4.8(1) implies that ∀a1 ,a2 ∈A a1 #W a2 , and (3) Theorem 4.8(2) implies that ∀b1 ,b2 ∈B b1 #W b2 . To prove that (A, B) ∈ YW , we need to show that it is not possible to have (A , B ) ∈ X such that A ⊆ A , B ⊆ B , and (A, B) = (A , B ) (i.e., A ⊂ A or B ⊂ B ). Suppose that A ⊂ A . There is an a ∈ T \ A such that ∀b∈B a →W b and ∀a∈A a#W a . Theorem 4.8(3) implies that a N ∩ a •
N

•

∩

N

•

b = ∅ for some b ∈ B. Let p ∈ a N ∩ a •

N

•

∩

N

•

b. Property 4.4

implies p = p. However, a ∈ A =N p and a ∈N p , and we find a contradiction (p = p • • and p = p). Suppose that B ⊂ B . There is a b ∈ T \ B such that ∀a∈A a →W b and ∀b∈B b#W b . Using Theorem 4.8(4) and Property 4.4, we can show that this leads to a contradiction. Therefore, (A, B) ∈ YW and pW ∈ PW . Let pw ∈ PW . We need to prove that there is a p ∈ P such that p N = pW •
NW N

•

p =

NW

•

pW and

•

. If pw = iw or pw = ow , then pw corresponds to i or o respectively. This

is a direct consequence of the construction given in Definition 4.9 involving iW , oW , TI , and TO . If pw ∈ {iw , ow }, then there are sets A and B such that (A, B) ∈ YW and pw = p(A,B) . such that
N α(N )

•

pw = A and pw

α(N )

•

= B. Remains to prove that there is a p ∈ P

•

p = A and p N = B. Since (A, B) ∈ YW implies that (A, B) ∈ XW , for any •

a ∈ A and b ∈ B there is a place connecting a and b (use a →W b and Theorem 4.1). 21

Using Theorem 4.8, we can prove that there is just one such place. Let p be this place.
• • • • Clearly, N p ⊆ A and p N ⊆ B. Remains to prove that N p = A and p N = B. Suppose that

a ∈N p \ A (i.e., N p = A). Select an arbitrary a ∈ A and b ∈ B. Using Theorem 4.6, we • • can show that a →W b. Using Theorem 4.8(1), we can show that a#W a . This holds for any a ∈ A and b ∈ B. Therefore, (A ∪ {a }, B) ∈ XW . However, this is not possible since (A, B) ∈ YW ((A, B) should be maximal). Therefore, we find a contradiction. We find a similar contradiction if we assume that there is a b ∈ p N \B. Therefore, we conclude • that N p = A and p N = B. • • ✷

B

A

C

D

E

Fig. 6. Another process model corresponding to the workflow log shown in Table 1.

Nets N1 , N2 and N5 shown in Figure 4 satisfy the requirements stated in Theorem 4.10. Therefore, it is no surprise that α is able to rediscover these nets. The net shown in Figure 1 is also an SWF-net with no short loops. Therefore, we can successfully rediscover the net if the AND-split and the AND-join are visible in the log. The latter assumption is not realistic if these two transitions do not correspond to real work. Given the fact the log shown in Table 1 does not list the occurrence of these events, indicates that this assumption is not valid. Therefore, the AND-split and the AND-join should be considered invisible. However, if we apply α to this log W = {ABCD, ACBD, AED}, then the result is quite surprising. The resulting net α(W ) is shown in Figure 6. To illustrate the α algorithm we show the result of each step using the log W = {ABCD, ACBD, AED} (i.e., a log like the one shown in Table 1): 1. TW = {A, B, C, D, E}, 2. TI = {A}, 3. TO = {D}, 22

4. XW = {({A}, {B}), ({A}, {C}), ({A}, {E}), ({B}, {D}), ({C}, {D}), ({E}, {D}), ({A}, {B, E}), ({A}, {C, E}), ({B, E}, {D}), ({C, E}, {D})}, 5. YW = {({A}, {B, E}), ({A}, {C, E}), ({B, E}, {D}), ({C, E}, {D})}, 6. PW = {iW , oW , p({A},{B,E}) , p{A},{C,E}) , p({B,E},{D}) , p({C,E},{D}) }, 7. FW = {(iW , A), (A, p({A},{B,E}) ), (p({A},{B,E}) , B) . . . , (D, oW )}, and 8. α(W ) = (PW , TW , FW ) (as shown in Figure 6). Although the resulting net is not an SWF-net it is a sound WF-net whose observable behavior is identical to the net shown in Figure 1. Also note that the WF-net shown in Figure 6 can be rediscovered although it is not an SWF-net. This example shows that the applicability is not limited to SWF-nets. However, for arbitrary sound WF-nets it is not possible to guarantee that they can be rediscovered. 4.4 Limitations of the α algorithm

As demonstrated through Theorem 4.10, the α algorithm is able to rediscover a large class of processes. However, we did not prove that the class of processes is maximal, i.e., that there is not a “better” algorithm able to rediscover even more processes. Therefore, we reflect on the requirements stated in Definition 4.3 (SWF-nets) and Theorem 4.10 (no short loops).

N6

A

D

C

B

E

N7

A

D

C

B

E

Fig. 7. The non-free-choice WF-net N6 cannot be rediscovered by the α algorithm.

23

C

N8

A D G E B F

C

N9

A D G E B F

Fig. 8. WF-net N8 cannot be rediscovered by the α algorithm. Nevertheless α returns a WF-net which is behavioral equivalent.

Let us first consider the requirements stated in Definition 4.3. To illustrate the necessity of the first two requirements consider figures 7 and 8. The WF-net N6 shown in Figure 7 is sound but not an SWF-net since the first requirement is violated (N6 is not free-choice). If we apply the mining algorithm to a complete workflow log W6 of N6 , we obtain the WF-nets N7 also shown in Figure 7 (i.e., α(W6 ) = N7 ). Clearly, N6 cannot be rediscovered using α. Although N7 is a sound SWF-net its behavior is different from N6 , e.g., workflow trace ACE is possible in N7 but not in N6 . This example motivates the first requirement in Definition 4.3. The second requirement is motivated by Figure 8. N8 violates the second requirement. If we apply the mining algorithm to a complete workflow log W8 of N8 , we obtain the WF-net α(W8 ) = N9 also shown in Figure 8. Although N9 is behaviorally equivalent, N8 cannot be rediscovered using the mining algorithm. Although the requirements stated in Definition 4.3 are necessary in order to prove that this class of workflow processes can be rediscovered on the basis of a complete workflow log, the applicability is not limited to SWF-nets. The examples given in this section show that in many situations a behaviorally equivalent WF-net can be derived. 24

Note that the third requirement stated in Definition 4.3 (no implicit places) can be dropped thus allowing even more behaviorally equivalent WF-nets. Even in the cases where the resulting WF-net is not behaviorally equivalent, the results are meaningful, e.g., the process represented by N7 is different from the process represented by N6 (cf. Figure 7). Nevertheless, N7 is similar and captures most of the behavior.
N10
A

B

C

E

D

N11
A

B

C

E

D

Fig. 9. Although both WF-nets are not behavioral equivalent they are identical with respect to >.

Another requirement imposed by Theorem 4.10 is the absence of short loops. We already showed examples motivating this requirement. However, it is clear that the α algorithm can be improved to deal with short loops. Unfortunately, this is less trivial than it may seem. We use Figure 9 and Figure 10 to illustrate this. Figure 9 shows two WF-nets. Although both WF-nets have clearly different behaviors their complete logs will have identical > relations. Note that the two WF-nets are not SWF-nets because of the non-free-choice construct involving E. However, the essence of the problem is the short loop involving C. Because of this short loop the α algorithm will create for both N10 and N11 the Petri net where C is unconnected to the rest of the net. Clearly, this can be improved since for complete logs of N10 and N11 the following relations hold: B > C, C > B, C > D, D > C, B C, C D, and B D. These relations suggest that B, C, and D are in parallel. However, in any complete log it can be seen that this is not the case since A is never followed by C. Despite this information, no algorithm will be able to distinguish N10 and N11 because they are identical with respect to >. Figure 10 gives another example demonstrating that dealing with short loops is far from trivial. N12 and N13 are SWF-nets that are behavioral equivalent, therefore, no algorithm will be able to 25

distinguish N12 from N13 (assuming the notion of completeness and not logging explicit start and end events). This problem may seem similar to Figure 8 or examples with and without implicit places. However, N12 and N13 are SWF-nets while N8 or examples with implicit places are just WF-nets not satisfying the requirements stated in Definition 4.3.
N12
A

B

D

C B

N13
D

A

C

Fig. 10. Both SWF-nets are behavioral equivalent and therefore any algorithm will be unable to distinguish N12 from N13 (assuming a notion of completeness based on >).

Despite the problems illustrated by Figure 9 and Figure 10, it is possible to solve the problem using a slightly stronger notion of completeness. Assume that it is possible to identify 1-loops and 2-loops. This can be done by looking for patterns AA (1-loop) and ABA (2-loop). Then refine relation > into a new relation >+ which splits the existing transitions involved in a short loop into 2 or 3 transitions. Tasks involved in a 1-loop are mapped onto three transitions (e.g., start, execute, and end). Tasks involved in a 2-loop but not a 1-loop are mapped onto two transitions (e.g., start and end). The goal of this translation is to make the short loops longer by using additional information about 1-loops and 2-loops (and while preserving the original properties). This refined relation >+ can be used as input for the α algorithm. This approach is able to deal with all possible situations except the one illustrated in Figure 10. Note that no algorithm will be able to distinguish the logs of the two processes shown in Figure 10. However, this is not a real problem since they are behaviorally equivalent. In formal terms: we can replace the requirement in Theorem 4.10 by the weaker requirement that there are no two transitions having identical input and output places. A detailed description of this preprocessing algorithm and a proof of its correctness are however beyond the scope of this paper. 26

Besides the explicit requirements stated in Definition 4.3 and Theorem 4.10, there are also a number of implicit requirements. As indicated before, hidden tasks cannot be detected (cf. the AND-split and AND-join in Figure 1). Moreover, our definition of a Petri net assumes that each transition bears a unique label. Instead, we could have used a labeled Petri net [39]. The latter choice would have been more realistic but also complicate matters enormously. The current definition of a WF-net assumes that each task appears only once in the network. Without this assumption, every occurrence of some task t could refer to one of multiple indistinguishable transitions with label t. Preliminary investigations show that the problems of “hidden tasks” and “duplicate tasks” are highly related, e.g., a WF-net with two transitions having the same label t has a corresponding WF-net with hidden transitions but only one transition labeled t. Moreover, there are relations between the explicit requirements stated in Definition 4.3 and Theorem 4.10 and the implicit requirements just mentioned, e.g., in some cases a non-free-choice WF-net can be made free choice by inserting hidden transitions. These findings indicate that the class of SWF-nets is close to the upper bound of workflow processes that can be mined successfully given the notion of completeness stated in Definition 3.5. To move beyond SWF-nets we either have to resort to heuristics or strengthen the notion of completeness (and thus require more observations). To conclude this section, we consider the complexity of the α algorithm. Event logs may be huge containing millions of events. Fortunately, the α algorithm is driven by relation >. The time it takes to build this relation is linear in the size of the log. The complexity of the remaining steps in α algorithm is exponential in the number of tasks. However, note that the number of tasks is typically less than 100 and does not depend on the size of the log. Therefore, the complexity is not a bottleneck for large-scale application.

5

Related Work

The idea of process mining is not new [7, 9–11, 19–24, 31–33, 41–45, 48–50]. Cook and Wolf have investigated similar issues in the context of software engineering processes. In [9] they describe three methods for process discovery: one using neural networks, one using a purely algorithmic approach, and one Markovian approach. The authors 27

consider the latter two the most promising approaches. The purely algorithmic approach builds a finite state machine where states are fused if their futures (in terms of possible behavior in the next k steps) are identical. The Markovian approach uses a mixture of algorithmic and statistical methods and is able to deal with noise. Note that the results presented in [9] are limited to sequential behavior. Related, but in a different domain, is the work presented in [29, 30] also using a Markovian approach restricted to sequential processes. Cook and Wolf extend their work to concurrent processes in [10]. They propose specific metrics (entropy, event type counts, periodicity, and causality) and use these metrics to discover models out of event streams. However, they do not provide an approach to generate explicit process models. Recall that the final goal of the approach presented in this paper is to find explicit representations for a broad range of process models, i.e., we want to be able to generate a concrete Petri net rather than a set of dependency relations between events. In [11] Cook and Wolf provide a measure to quantify discrepancies between a process model and the actual behavior as registered using event-based data. The idea of applying process mining in the context of workflow management was first introduced in [7]. This work is based on workflow graphs, which are inspired by workflow products such as IBM MQSeries workflow (formerly known as Flowmark) and InConcert. In this paper, two problems are defined. The first problem is to find a workflow graph generating events appearing in a given workflow log. The second problem is to find the definitions of edge conditions. A concrete algorithm is given for tackling the first problem. The approach is quite different from other approaches: Because the nature of workflow graphs there is no need to identify the nature (AND or OR) of joins and splits. As shown in [27], workflow graphs use true and false tokens which do not allow for cyclic graphs. Nevertheless, [7] partially deals with iteration by enumerating all occurrences of a given task and then folding the graph. However, the resulting conformal graph is not a complete model. In [33], a tool based on these algorithms is presented. Schimm [41, 42, 45] has developed a mining tool suitable for discovering hierarchically structured workflow processes. This requires all splits and joins to be balanced. Herbst and Karagiannis also address the issue of process mining in the context of workflow management [21, 19, 20, 23, 24, 22] using an inductive approach. The work presented in [22, 24] is limited to sequential models. The approach described in [21, 28

19, 20, 23] also allows for concurrency. It uses stochastic task graphs as an intermediate representation and it generates a workflow model described in the ADONIS modeling language. In the induction step task nodes are merged and split in order to discover the underlying process. A notable difference with other approaches is that the same task can appear multiple times in the workflow model, i.e., the approach allows for duplicate tasks. The graph generation technique is similar to the approach of [7, 33]. The nature of splits and joins (i.e., AND or OR) is discovered in the transformation step, where the stochastic task graph is transformed into an ADONIS workflow model with blockstructured splits and joins. In contrast to the previous papers, our work [31, 32, 48–50] is characterized by the focus on workflow processes with concurrent behavior (rather than adding ad-hoc mechanisms to capture parallelism). In [48–50] a heuristic approach using rather simple metrics is used to construct so-called “dependency/frequency tables” and “dependency/frequency graphs”. In [31] another variant of this technique is presented using examples from the health-care domain. The preliminary results presented in [31, 48–50] only provide heuristics and focus on issues such as noise. The approach described in this paper differs from these approaches in the sense that for the α algorithm it is proven that for certain subclasses it is possible to find the right workflow model. In [4] the EMiT tool is presented which uses an extended version of α algorithm to incorporate timing information. Note that in this paper there is no detailed description of the α algorithm nor a proof of its correctness. Process mining can be seen as a tool in the context of Business (Process) Intelligence (BPI). In [18, 40] a BPI toolset on top of HP’s Process Manager is described. The BPI tools set includes a so-called “BPI Process Mining Engine”. However, this engine does not provide any techniques as discussed before. Instead it uses generic mining tools such as SAS Enterprise Miner for the generation of decision trees relating attributes of cases to information about execution paths (e.g., duration). In order to do workflow mining it is convenient to have a so-called “process data warehouse” to store audit trails. Such as data warehouse simplifies and speeds up the queries needed to derive causal relations. In [13, 34–36] the design of such warehouse and related issues are discussed in the context of workflow logs. Moreover, [36] describes the PISA tool which can be used to extract performance metrics from workflow logs. Similar diagnostics are provided by the ARIS 29

Process Performance Manager (PPM) [25]. The later tool is commercially available and a customized version of PPM is the Staffware Process Monitor (SPM) [46] which is tailored towards mining Staffware logs. Note that none of the latter tools is extracting the process model. The main focus is on clustering and performance analysis rather than causal relations as in [7, 9–11, 19–24, 31–33, 41–45, 48–50]. More from a theoretical point of view, the rediscovery problem discussed in this paper is related to the work discussed in [8, 16, 17, 38]. In these papers the limits of inductive inference are explored. For example, in [17] it is shown that the computational problem of finding a minimum finite-state acceptor compatible with given data is NP-hard. Several of the more generic concepts discussed in these papers could be translated to the domain of process mining. It is possible to interpret the problem described in this paper as an inductive inference problem specified in terms of rules, a hypothesis space, examples, and criteria for successful inference. The comparison with literature in this domain raises interesting questions for process mining, e.g., how to deal with negative examples (i.e., suppose that besides log W there is a log V of traces that are not possible, e.g., added by a domain expert). However, despite the many relations with the work described in [8, 16, 17, 38] there are also many differences, e.g., we are mining at the net level rather than sequential or lower level representations (e.g., Markov chains, finite state machines, or regular expressions). Additional related work is the seminal work on regions [14]. This work investigates which transition systems can be represented by (compact) Petri nets (i.e., the so-called synthesis problem). Although the setting is different and our notion of completeness is much weaker than knowing the transition system, there are related problems such as duplicate transitions, etc.

6

Conclusion

In this paper we addressed the workflow rediscovery problem. This problem was formulated as follows: “Find a mining algorithm able to rediscover a large class of sound WF-nets on the basis of complete workflow logs.” We presented the α algorithm that is able to rediscover a large and relevant class of workflow processes (SWF-nets). Through examples we also showed that the algorithm provides interesting analysis results for 30

workflow processes outside this class. At this point in time, we are improving the mining algorithm such that it is able to rediscover an even larger class of WF-nets. We have tackled the problem of short loops and are now focusing on hidden tasks, duplicate tasks, and advanced routing constructs. However, given the observation that the class of SWF-nets is close to the upper limit of what one can do assuming this notion of completeness, new results will either provide heuristics or require stronger notions of completeness (i.e., more observations). It is important to see the results presented in this paper in the context of a larger effort [31, 32, 48–50]. The rediscovery problem is not a goal by itself. The overall goal is to be able to analyze any workflow log without any knowledge of the underlying process and in the presence of noise. The theoretical results presented in this paper provide insights that are consistent with empirical results found earlier [31, 32, 48–50]. It is quite interesting to see that the challenges encountered in practice match the challenges encountered in theory. For example, the fact that workflow process exhibiting non-free-choice behavior (i.e., violating the first requirement of Definition 4.3) are difficult to mine was observed both in theory and in practice. Therefore, we consider the work presented in this paper as a stepping stone for good and robust process mining techniques. We have applied our workflow mining techniques to two real applications. The first application is in health-care where the flow of multi-disciplinary patients is analyzed. We have analyzed workflow logs (visits to different specialists) of patients with peripheral arterial vascular diseases of the Elizabeth Hospital in Tilburg and the Academic Hospital in Maastricht. Patients with peripheral arterial vascular diseases are a typical example of multi-disciplinary patients. We have preliminary results showing that process mining is very difficult given the “spaghetti-like” nature of this process. Only by focusing on specific tasks and abstracting from infrequent tasks we are able to successfully mine such processes. The second application concerns the processing of fines by the CJIB (Centraal Justitieel Incasso Bureau), the Dutch Judicial Collection Agency located in Leeuwarden. We have successfully mined the process using information of 130136 cases. The process comprises 99 tasks and has been validated by the CJIB. This positive result shows that process mining based on the α algorithm and using tools like EMiT and Little Thumb 31

is feasible for at least structured processes. These findings are encouraging and show the potential of the α algorithm presented in this paper.

Acknowledgements The authors would like to thank Ana Karla Alves de Medeiros and Eric Verbeek for proof-reading earlier versions of this paper and Boudewijn van Dongen for his efforts in developing EMiT and solving the problem of short loops.

References
1. W.M.P. van der Aalst. Verification of Workflow Nets. In P. Az´ma and G. Balbo, editors, Application and e Theory of Petri Nets 1997, volume 1248 of Lecture Notes in Computer Science, pages 407–426. SpringerVerlag, Berlin, 1997. 2. W.M.P. van der Aalst. The Application of Petri Nets to Workflow Management. The Journal of Circuits, Systems and Computers, 8(1):21–66, 1998. 3. W.M.P. van der Aalst, J. Desel, and A. Oberweis, editors. Business Process Management: Models, Techniques, and Empirical Studies, volume 1806 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, 2000. 4. W.M.P. van der Aalst and B.F. van Dongen. Discovering Workflow Performance Models from Timed Logs. In Y. Han, S. Tai, and D. Wikarski, editors, International Conference on Engineering and Deployment of Cooperative Information Systems (EDCIS 2002), volume 2480 of Lecture Notes in Computer Science, pages 45–63. Springer-Verlag, Berlin, 2002. 5. W.M.P. van der Aalst and K.M. van Hee. Workflow Management: Models, Methods, and Systems. MIT press, Cambridge, MA, 2002. 6. W.M.P. van der Aalst, A.J.M.M. Weijters, and L. Maruster. Workflow Mining: Which Processes can be Rediscovered? BETA Working Paper Series, WP 74, Eindhoven University of Technology, Eindhoven, 2002. 7. R. Agrawal, D. Gunopulos, and F. Leymann. Mining Process Models from Workflow Logs. In Sixth International Conference on Extending Database Technology, pages 469–483, 1998. 8. D. Angluin and C.H. Smith. Inductive Inference: Theory and Methods. Computing Surveys, 15(3):237–269, 1983. 9. J.E. Cook and A.L. Wolf. Discovering Models of Software Processes from Event-Based Data. ACM Transactions on Software Engineering and Methodology, 7(3):215–249, 1998. 10. J.E. Cook and A.L. Wolf. Event-Based Detection of Concurrency. In Proceedings of the Sixth International Symposium on the Foundations of Software Engineering (FSE-6), pages 35–45, 1998. 11. J.E. Cook and A.L. Wolf. Software Process Validation: Quantitatively Measuring the Correspondence of a Process to a Model. ACM Transactions on Software Engineering and Methodology, 8(2):147–176, 1999.

32

12. J. Desel and J. Esparza. Free Choice Petri Nets, volume 40 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, Cambridge, UK, 1995. 13. J. Eder, G.E. Olivotto, and Wolfgang Gruber. A Data Warehouse for Workflow Logs. In Y. Han, S. Tai, and D. Wikarski, editors, International Conference on Engineering and Deployment of Cooperative Information Systems (EDCIS 2002), volume 2480 of Lecture Notes in Computer Science, pages 1–15. Springer-Verlag, Berlin, 2002. 14. A. Ehrenfeucht and G. Rozenberg. Partial (Set) 2-Structures - Part 1 and Part 2. Acta Informatica, 27(4):315–368, 1989. 15. L. Fischer, editor. Workflow Handbook 2001, Workflow Management Coalition. Future Strategies, Lighthouse Point, Florida, 2001. 16. E.M. Gold. Language Identfication in the Limit. Information and Control, 10(5):447–474, 1967. 17. E.M. Gold. Complexity of Automaton Identification from Given Data. Information and Control, 37(3):302– 320, 1978. 18. D. Grigori, F. Casati, U. Dayal, and M.C. Shan. Improving Business Process Quality through Exception Understanding, Prediction, and Prevention. In P. Apers, P. Atzeni, S. Ceri, S. Paraboschi, K. Ramamohanarao, and R. Snodgrass, editors, Proceedings of 27th International Conference on Very Large Data Bases (VLDB’01), pages 159–168. Morgan Kaufmann, 2001. 19. J. Herbst. A Machine Learning Approach to Workflow Management. In Proceedings 11th European Conference on Machine Learning, volume 1810 of Lecture Notes in Computer Science, pages 183–194. SpringerVerlag, Berlin, 2000. 20. J. Herbst. Dealing with Concurrency in Workflow Induction. In U. Baake, R. Zobel, and M. Al-Akaidi, editors, European Concurrent Engineering Conference. SCS Europe, 2000. 21. J. Herbst. Ein induktiver Ansatz zur Akquisition und Adaption von Workflow-Modellen. PhD thesis, Universit¨t Ulm, November 2001. a 22. J. Herbst and D. Karagiannis. Integrating Machine Learning and Workflow Management to Support Acquisition and Adaptation of Workflow Models. In Proceedings of the Ninth International Workshop on Database and Expert Systems Applications, pages 745–752. IEEE, 1998. 23. J. Herbst and D. Karagiannis. An Inductive Approach to the Acquisition and Adaptation of Workflow Models. In M. Ibrahim and B. Drabble, editors, Proceedings of the IJCAI’99 Workshop on Intelligent Workflow and Process Management: The New Frontier for AI in Business, pages 52–57, Stockholm, Sweden, August 1999. 24. J. Herbst and D. Karagiannis. Integrating Machine Learning and Workflow Management to Support Acquisition and Adaptation of Workflow Models. International Journal of Intelligent Systems in Accounting, Finance and Management, 9:67–92, 2000. 25. IDS Scheer. ARIS Process Performance Manager (ARIS PPM). http://www.ids-scheer.com, 2002.

33

26. S. Jablonski and C. Bussler. Workflow Management: Modeling Concepts, Architecture, and Implementation. International Thomson Computer Press, London, UK, 1996. 27. B. Kiepuszewski. flows. Expressiveness and Suitability of Languages for Control Flow Modelling in WorkAvailable via

PhD thesis, Queensland University of Technology, Brisbane, Australia, 2002.

http://www.tm.tue.nl/it/research/patterns. 28. F. Leymann and D. Roller. Production Workflow: Concepts and Techniques. Prentice-Hall PTR, Upper Saddle River, New Jersey, USA, 1999. 29. H. Mannila and D. Rusakov. Decomposing Event Sequences into Independent Components. In V. Kumar and R. Grossman, editors, Proceedings of the First SIAM Conference on Data Mining, pages 1–17. SIAM, 2001. 30. H. Mannila, H. Toivonen, and A.I. Verkamo. Discovery of Frequent Episodes in Event Sequences. Data Mining and Knowledge Discovery, 1(3):259–289, 1997. 31. L. Maruster, W.M.P. van der Aalst, A.J.M.M. Weijters, A. van den Bosch, and W. Daelemans. Automated Discovery of Workflow Models from Hospital Data. In B. Kr¨se, M. de Rijke, G. Schreiber, and M. o van Someren, editors, Proceedings of the 13th Belgium-Netherlands Conference on Artificial Intelligence (BNAIC 2001), pages 183–190, 2001. 32. L. Maruster, A.J.M.M. Weijters, W.M.P. van der Aalst, and A. van den Bosch. Process Mining: Discovering Direct Successors in Process Logs. In Proceedings of the 5th International Conference on Discovery Science (Discovery Science 2002), volume 2534 of Lecture Notes in Artificial Intelligence, pages 364–373. SpringerVerlag, Berlin, 2002. 33. M.K. Maxeiner, K. K¨ spert, and F. Leymann. Data Mining von Workflow-Protokollen zur teilautomau tisierten Konstruktion von Prozemodellen. In Proceedings of Datenbanksysteme in B¨ro, Technik und u Wissenschaft, pages 75–84. Informatik Aktuell Springer, Berlin, Germany, 2001. 34. M. zur M¨hlen. Process-driven Management Information Systems u Combining Data Warehouses and

Workflow Technology. In B. Gavish, editor, Proceedings of the International Conference on Electronic Commerce Research (ICECR-4), pages 550–566. IEEE Computer Society Press, Los Alamitos, California, 2001. 35. M. zur M¨hlen. Workflow-based Process Controlling-Or: What You Can Measure You Can Control. In u L. Fischer, editor, Workflow Handbook 2001, Workflow Management Coalition, pages 61–77. Future Strategies, Lighthouse Point, Florida, 2001. 36. M. zur M¨hlen and M. Rosemann. Workflow-based Process Monitoring and Controlling - Technical and u Organizational Issues. In R. Sprague, editor, Proceedings of the 33rd Hawaii International Conference on System Science (HICSS-33), pages 1–10. IEEE Computer Society Press, Los Alamitos, California, 2000. 37. T. Murata. Petri Nets: Properties, Analysis and Applications. Proceedings of the IEEE, 77(4):541–580, April 1989.

34

38. L. Pitt. Inductive Inference, DFAs, and Computational Complexity. In K.P. Jantke, editor, Proceedings of International Workshop on Analogical and Inductive Inference (AII), volume 397 of Lecture Notes in Computer Science, pages 18–44. Springer-Verlag, Berlin, 1889. 39. W. Reisig and G. Rozenberg, editors. Lectures on Petri Nets I: Basic Models, volume 1491 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, 1998. 40. M. Sayal, F. Casati, and M.C. Shan U. Dayal. Business Process Cockpit. In Proceedings of 28th International Conference on Very Large Data Bases (VLDB’02), pages 880–883. Morgan Kaufmann, 2002. 41. G. Schimm. Process Mining. http://www.processmining.de/. 42. G. Schimm. Generic Linear Business Process Modeling. In S.W. Liddle, H.C. Mayr, and B. Thalheim, editors, Proceedings of the ER 2000 Workshop on Conceptual Approaches for E-Business and The World Wide Web and Conceptual Modeling, volume 1921 of Lecture Notes in Computer Science, pages 31–39. Springer-Verlag, Berlin, 2000. 43. G. Schimm. Process Mining elektronischer Gesch¨ftsprozesse. a In Proceedings Elektronische

Gesch¨ftsprozesse, 2001. a 44. G. Schimm. Process Mining linearer Prozessmodelle - Ein Ansatz zur automatisierten Akquisition von Prozesswissen. In Proceedings 1. Konferenz Professionelles Wissensmanagement, 2001. 45. G. Schimm. Process Miner - A Tool for Mining Process Schemes from Event-based Data. In S. Flesca and G. Ianni, editors, Proceedings of the 8th European Conference on Artificial Intelligence (JELIA), volume 2424 of Lecture Notes in Computer Science, pages 525–528. Springer-Verlag, Berlin, 2002. 46. Staffware. Staffware Process Monitor (SPM). http://www.staffware.com, 2002. 47. H.M.W. Verbeek, T. Basten, and W.M.P. van der Aalst. Diagnosing Workflow Processes using Woflan. The Computer Journal, 44(4):246–279, 2001. 48. A.J.M.M. Weijters and W.M.P. van der Aalst. Process Mining: Discovering Workflow Models from EventBased Data. In B. Kr¨se, M. de Rijke, G. Schreiber, and M. van Someren, editors, Proceedings of the 13th o Belgium-Netherlands Conference on Artificial Intelligence (BNAIC 2001), pages 283–290, 2001. 49. A.J.M.M. Weijters and W.M.P. van der Aalst. Rediscovering Workflow Models from Event-Based Data. In V. Hoste and G. de Pauw, editors, Proceedings of the 11th Dutch-Belgian Conference on Machine Learning (Benelearn 2001), pages 93–100, 2001. 50. A.J.M.M. Weijters and W.M.P. van der Aalst. Workflow Mining: Discovering Workflow Models from EventBased Data. In C. Dousson, F. H¨ppner, and R. Quiniou, editors, Proceedings of the ECAI Workshop on o Knowledge Discovery and Spatial Data, pages 78–84, 2002.

35

About the authors

Wil van der Aalst is a full professor of Information Systems and head of the section of Information and Technology of the Department of Technology Management at Eindhoven University of Technology. He is also a part-time full professor at the Computing Science faculty at the department of Mathematics and Computer Science at the same university and an adjunct professor at Queensland University of Technology. His research interests include information systems, simulation, Petri nets, process models, workflow management systems, verification techniques, enterprise resource planning systems, computer supported cooperative work, and interorganizational business processes.

Ton Weijters is associate professor at the Department of Technology Management of the Eindhoven University of Technology (TUE), and member of the BETA research group. Currently he is working on (i) the application of Knowledge Engineering and Machine Learning techniques for planning, scheduling, and process mining (ii) fundamental research in the domain of Machine Learning and Knowledge Discovering. He is the author of many scientific publications in the mentioned research field.

Laura Maruster received her B.S. degree in 1994 and M.S. in 1995, both in Computer Science Department at West University of Timisoara, Romania. At present she is a Ph.D candidate of the Department of Technology Management of Eindhoven University of Technology, Eindhoven, The Netherlands. Her research interests include induction of machine learning and statistical models, process mining and knowledge discovery.

Contact information
Corresponding Author Prof.dr.ir. W.M.P. van der Aalst Eindhoven University of Technology Faculty of Technology and Management (PAV D2) Department of Information and Technology PO Box 513 36

NL-5600 MB Eindhoven The Netherlands

Phone: +31 40 247.4295/2290 Fax: +31 40 243.2612 E-mail: [email protected] WWW: http://www.tm.tue.nl/it/staff/wvdaalst/

Author 2 Dr.ir. A.J.M.M. Weijters Eindhoven University of Technology Faculty of Technology and Management Department of Information and Technology PO Box 513 NL-5600 MB Eindhoven The Netherlands

Phone: +31 40 247.3857/2290 Fax: +31 40 243.2612 E-mail: [email protected] WWW: http://www.tm.tue.nl/it/staff/tweijters

Author 3 M.Sc. L. Maruster Eindhoven University of Technology Faculty of Technology and Management Department of Information and Technology PO Box 513 NL-5600 MB Eindhoven The Netherlands 37

Phone: +31 40 247.3703/2290 Fax: +31 40 243.2612 E-mail: [email protected] WWW: http://www.tm.tue.nl/it/staff/lmaruster

38