Process mining allows for a fact-based analysis of business processes by discovering descriptive process models. However, providing an overall view of the process is not enough. To gain valuable insights and identify potentials for process improvement, it is crucial to lter the underlying event log to a subset of cases of interest. Usually this is done by attribute-based lters, e.g. ltering for a speci c vendor, production line or time period. However, sometimes the interesting cases are de ned by a more complex pattern, e.g. by a sequence of certain events. In this demo, we present a new feature of Celonis that allows the analysts to lter for such complex patterns by de ning a regular expression. Due to its integration into the Celonis Process Query Language, it can be applied to arbitrary analysis components. This allows for ad-hoc ltering by experienced analysts as well as pre-de ned lters for business users.
Process mining provides business analysts and process owners a fact-based view
on their processes, e.g. by discovering a descriptive process model [
Usually, these drill-downs are performed along dimensions stored with the data (e.g. region, vendor, etc.) or based on KPIs like throughput time. However, all of them are based on simple attribute values but do not consider the process behavior, such as the ow of activities. For example, a user might be interested in cases where a delivery block has been set but has never been removed afterwards.
In this demo, we present a novel feature of Celonis Process Mining which allows to lter cases based on user-de ned process ow patterns. These patterns are de ned as regular expressions that are processed by a speci c lter operator
REGULAR EXPRESSION
SEARCH
INSIGHTS USER REPLAY
PROCESS MAP TRANSLATION
DFA
FILTER
EVENT LOG
DRILL-DOWN integrated into the Celonis Process Query Language (PQL). This allows for applying the lter to a speci c analysis component (e.g. a table or chart) or even the entire analysis. As shown in this demo, advanced users can also interactively lter the cases by entering a regular expression describing the process ow pattern of interest. This enables them to dynamically explore the processes in depth in order to nd potential aws or undesired behaviors in the process.
In the remainder of this paper, we provide an overview of the pattern-based case lters and its capabilities in Section 2. Section 3 discusses related work and Section 4 describes the presented demo and its setting. 2
The searched behavior is de ned as a regular expression { a well-known concept for pattern-matching in programming. In contrast to programming, we do not apply them to strings but to activity sequences. The syntax is inspired by the widely used Perl syntax but has some adoptions to its application to event logs.
Figure 1 gives an overview of our approach. The user-de ned regular
expression is translated into a deterministic nite automaton (DFA) by applying the
Berry-Sethi-algorithm [
To use the pattern-based case lters in an analysis, the Celonis PQL provides the match process regex(column, pattern ) operator. The rst parameter is the activity column in the data model which stores the executed activity for each event in the event log. The second parameter is the regular expression describing the searched pattern. The result of the operator is a new integer column attached to the case table. For each case matching the pattern, the value is 1 while for non-matching cases the value is 0. This allows for the seamless integration of the operator into any PQL lter formula de ned in the analysis. Binding the regular expression to a variable allows to dynamically adjust the pattern by selecting pre-de ned patterns or by editing it in an ad-hoc fashion in a text- eld. The following language constructs can be used to de ne regular expressions which can also be nested to de ne complex patterns.
Activities: Activities are the primitives of the regular expressions and are identi ed by their single-quoted name (e.g. ’Create Invoice’). If the activity should match multiple activities with similar name, it is also possible to use wildcard matching for activities (e.g. LIKE ’Create%’). Instead of the activity name, the ANY keyword can be used to match any arbitrary activity. Sequences: A sequence de nes a directly follows relationship between two regular expressions. Usually, this is expressed by concatenating two symbols of the regular expression. However, as we have activity names as primitives, this would result in a very confusing syntax. Therefore, we use >> to express a sequence of two activities (e.g., ’Create Invoice’ >> ’Clear Invoice’). Choices: To choose between multiple valid regular expressions, we separate the options with | (e.g. ’Change Quantity’ | ’Change Price’). Alternatively, we can also give a comma-separated list of activity names surrounded by squared brackets to de ne a choice (e.g. [’Change Quantity’, ’Change Price’]). While the former is applicable to any regular expression, the latter only accepts activities as the primitives. Though the latter allows to invert the set of activities. For example, [ ! ’Change Quantity’, ’Change Price’] matches all activities except Change Quantity and Change Price. Quanti ers: With quanti ers ( , +, ?) one can declare that a regular expression should not match only once, but with a certain cardinality. A regular expression must be surrounded by brackets when applying a quanti er to it. While (’Create Invoice’) matches any arbitrary number of occurrence, (’Create Invoice’)+ requires the activity Create Invoice to occur at least once. With (’Create Invoice’)? one can mark it as optional, i.e. Create Invoice occurs once or not.
Match at start/end: To declare that a regular expression must match at the start (i.e. from the rst event of the trace), we can prepend a circum ex (^) to it. Analogously, we can append $ to the regular expression if it must match at the end. If the regular expression is not marked to match at the start / end, we implicitly prepend / append (ANY) to the regular expression for convenience in order to match the pattern anywhere within a trace. Please note that this may interfere with other quanti ers. For example, the regular expression (’Create Invoice’)? will also match traces having two or more consecutive Create Invoice activities, as the additional activities are consumed by the implicitly added (ANY) .
Due to the usage of activities as primitives, regular expressions may become quite long. To reduce the user's e ort to construct complex expressions and to improve their readability, regular expressions can be assigned to an alias which then can be referenced in other regular expressions. Note that the regular expressions must be comma-separated and the last regular expression must not have an alias, e.g. ’Change Quantity’ | ’Change Price’ AS A, ’Create Invoice’ AS B, A >> (ANY) >> A >> B
Instead of the activity column, it is also possible to use any other string column of the activity table. This enables the user to de ne search patterns not only over the activities but also over other event attributes like the resource. Due to its integration into PQL, also columns of other type can be converted into strings so e ectively each column of the activity table can be used.
The pattern-based case lters are available in the Celonis Intelligent Business Cloud (IBC) and in Celonis Process Mining 4.5. It is accessible to thousands of users and actively used by various Celonis customers.
The Celonis Academic Alliance1 o ers free licenses for academic purposes. 3
Filtering event logs by a speci c pattern of behavior can also be achieved by
other approaches. For example, LTL Checker [
Similar results can be achieved by drawing the behavioral pattern as
process model, then checking its conformance [
In this demo, we show a ready-to-use analysis running in the Celonis IBC. The components of the analysis are con gured to apply a lter containing the process match regex operator. The regular expressions used in the component lters are bound to a global variable to share the same regular expression among di erent components. To change this variable one can enter a new regular expression into one of the text elds. Alternatively, it is also possible to select a pre-de ned regular expression from the drop-down menu on the left-hand side of the analysis. This shows that the new lter is capable to be used either by power users that interactively lter the cases by de ning a regular expression ad-hoc or by less experienced users who select some prepared lters from a menu. 1 https://www.celonis.com/academic-signup
For this demo, we use a demo data set of a standard order-to-cash (O2C) process. The data consists of more than 6,000,000 events of almost 1,000,000 cases with almost 500 di erent traces.
Figure 2 shows a screenshot of the demo. The process map is ltered to all cases with an unresolved delivery block. This is done by the regular expression ’Set Delivery Block’ >> ([! ’Release Delivery’]) $ entered in the text eld below. As indicated by the numbers on the left-hand side, the ltered data comprises 8,605 cases forming 53 unique variants (traces). A screencast showing a short walk-through of the demo is available online2.