-Process analysts use process mining techniques to obtain fact-based knowledge from event logs about how business processes are actually executed in organizations. Often process discovery is the first step in their analytical workflow. However, when working with large amount of data and complex processes, exploring as-is process models to obtain interesting and insightful knowledge can be challenging. We propose PROCESSEXPLORER, an interactive visual recommendation system for process discovery to facilitate event log exploration. PROCESSEXPLORER automatically analyzes the event log to obtain promising subsets of cases, evaluates interesting process performance indicators, and recommends those that are most interesting and insightful. Our system uses multi-perspective trace clustering to identify candidate cases of interest and a deviation-based approach to assess the interestingness of process performance indicators. We implemented PROCESSEXPLORER as a standalone desktop application that allows to explore any process and any event log. Our demo shows how the workflow of analysts is supported by the system through suggesting subset and insights recommendations.
Index Terms—process discovery, variants analysis, log preprocessing, trace clustering, statistical hypothesis testing
Nowadays, information systems in organizations support and automate the processing of business transactions. These systems are typically integrated into companies’ business processes and record the activities that have been executed in the form of an event log. Process mining aims at providing an accurate view of how processes are actually executed in organizations. In particular, process discovery reconstructs asis process models from event logs which can be used for further analysis. A wide range of process mining tools has been established that implement process discovery and analysis methods to support analysts to obtain valuable knowledge. With this knowledge, process issues can be identified and optimizations can be implemented.
In this paper, we introduce the PROCESSEXPLORER system which provides recommendations to the analyst on how to select a subset of cases and what statistics may be interesting and insightful. Our system is inspired by the workflow that analysts typically perform when working with process mining tools. The visual inspection of the discovered process model is the initial starting point of any process mining project. Due to the massive growth of data, the increasing process complexity, and the flexible execution of business processes in organizations, visual exploration and analysis are getting more and more challenging. Often the analyst is confronted with a spaghetti-like process map which by itself does not necessarily lead to useful insights. Without extensive knowledge about the underlying process, selecting the right set of cases to find interesting and valuable insights or trends is non-trivial. In current process mining tools, most of these analysis steps are performed manually, leading to a lot of repetitive work which hampers efficient exploration and analysis.
PROCESSEXPLORER extends the interactive visual
exploration capabilities in today’s process mining tools by providing
automatic guidance to the analyst. Our tool integrates several
recommendation suggestions in a user-friendly manner to
improve overall process discovery exploration:
1) Subset Recommendation. PROCESSEXPLORER
recommends subsets of interesting cases to allow analysts
quickly inspect the different process behaviors observed
in the event log. Different from the manual filtering
that requires expert knowledge, subset recommendations
are automatically derived by mining process behavior
patterns from the dataset to simplify subset selection.
2) Insights Recommendation. After selecting a subset of
cases, PROCESSEXPLORER automatically computes a
range of relevant process performance indicators to
show interesting deviations. Analysts are guided towards
interesting statistics that they usually would compute
manually.
3) Recommendation Ranking. In order to prevent the
analyst from inspecting only a limited subset of cases,
PROCESSEXPLORER provides the analyst with the most
diversifying recommendations by applying diversifying
top-k ranking [
PROCESSEXPLORER is agnostic to the process and event log that is being analyzed. Any process and any event log in the standardized IEEE XES format can be used. Furthermore, the analyst does not need to setup any configuration or specify parameter values. Prior knowledge about the process or the event log is not required. PROCESSEXPLORER obtains all the necessary information from the event log itself.
We used PROCESSEXPLORER in a case study on the BPI
Challenge 2019 event log collected from a large company
to investigate the procurement handling process [
PROCESSEXPLORER extends process mining tools by introducing a recommendation engine to support analysts selecting interesting subsets of cases and generating insightful statistics. In particular, our system allows to quickly scan unknown processes in event logs to obtain knowledge about how the process is actually executed and where potential issues can be found. PROCESSEXPLORER provides two types of recommendations and a ranking mechanism.
The first type of recommendation suggests subsets of cases
that contain interesting process behavior patterns. We are
particularly interested in patterns that combine the control
flow and the data perspective. This is inspired by the manual
work of analysts who not only filter cases by the sequence
of activities but also by attributes. This is often used to
compare different departments, products, or company
locations. To support analysts during the selection of appropriate
subsets of cases, PROCESSEXPLORER automatically analyzes
the given event log to find such patterns using trace clustering.
Specifically, we apply multi-perspective trace clustering [
Another typical task in process mining is to investigate and
compare a range of process performance indicators (PPIs),
such as the number of activities, the total duration time,
the duration time between activities, the directly
followedby relation, and the existence of activities. These are either
directly visualized in the process map or separately displayed
in the form of statistical charts or single values. Existing
process mining tools provide assistance by offering the
possibility to create dashboards with predefined PPIs which will
update immediately if a different case selection is made.
Still, each PPI needs to be investigated one after another to
identify deviations which is time-consuming and error-prone.
PROCESSEXPLORER automatically computes these PPIs for a
selected subset and identifies those ones that may be
interesting to the user by performing statistical significance testing.
Compared to dashboards that are static with respect to the
computed PPIs, PROCESSEXPLORER reevaluates the PPIs for
each applied subset recommendation individually. Only PPIs
that are significantly different from the rest of the cases in the
event log are considered as an interesting insight [
Lastly, PROCESSEXPLORER ranks the recommendations
based on the interestingness score [
During our experiments, we found that certain insights co-occur with each other which unnecessarily increases the number of insights recommendations. PROCESSEXPLORER clusters similar insights recommendations using the Spearman’s rank-order correlation.
Subset recommendations are assigned a score based on the
insights scores and the number of cases that are contained
in the subset. We obtain the top-k subset recommendations
using the top-k diversifying ranking algorithm [
PROCESSEXPLORER is a standalone interactive process
mining tool to demonstrate the proposed guidance
capabilities. As mentioned earlier, it allows importing any
standardized IEEE XES event log and works without specifying
any additional parameter value. We give a walk-through of
PROCESSEXPLORER by inspecting the procurement handling
process of the BPI Challenge 2019 event log [
a) Process Map: The most prominent component in PROCESSEXPLORER is the process map. It visualizes the activities and transitions that have been observed in the event log. Activities and transitions can be filtered by their relative occurrence using the slider at the bottom right. Figure 1 shows the process map of a selected subset recommendation.
ranked list of subset recommendations is shown. Subset recommendations can be modified and adjusted by the user, enabling to further refine the selection of cases interactively. Users can add a happy path filter, a variant filter, a start and end activity filter, and an activity occurrence filter. Figure 1 shows the 8 subset recommendations that are suggested for the currently selected subset of cases.
c) Subset Statistics: On the lower right side, basic statistics of the selected subset recommendation are shown which give an overview of the cases in the subset. The statistics show how the subset selection compares to the original event log and highlights the event distribution, the variant distribution, and the number of selected cases. Based on the statistics, the user can decide which subset recommendation to apply. In the example, the selected subset recommendation selects 6 events, and 1 out of 4 variants.
PROCESSEXPLORER shows the insights recommendations for the current subset. Insights recommendations are automatically updated each time the subset of cases is modified. The system computes a range of basic PPIs which are typically analyzed by users. We distinguish between case- and subsetbased insights. Depending on the insight type, a different visualization is shown to the user. Figure 1 shows a portion of the obtained insights recommendations. For instance, the first insight refers to the directly followed-by relation between the “Record Invoice Receipt” and “Remove Payment Block” activities, which occurs more often in the applied subset. Furthermore, we can see that the activity “Receive Order Confirmation” is mostly executed by “user 029”.
e) Stage Views: For easier navigation between the different subset recommendations, PROCESSEXPLORER introduces stage views. Each time the user decides to apply a subset recommendation a new stage view is generated. A stage view stores the selected cases and the computed insight recommendations. Stages are organized as a hierarchical structure such that each refinement of a selection results in a new hierarchy level. For each stage view, subset and insights recommendations are computed, so recommendations can be successively refined.
PROCESSEXPLORER is built of three main components: the event log manager (XLogManager), the stage manager (XStageManager), and the recommendation manager (RecommendationManager). All three components are open for extension, such that other event log formats, stage management capabilities, subset and insights recommendation approaches can be integrated. Figure 2 shows the overall architecture of PROCESSEXPLORER.
Event logs are imported as an OpenXES XLog object and stored in-memory using the XESlite extension. Each loaded log is stored in the XLogData object structure which links to the XLog object and stores the basic statistics of the log. The XStageManager is responsible for managing the views of PROCESSEXPLORER, storing a history of all stages visited by the user. For an active stage, the XStageManager retrieves the recommendations from the RecommendationManager which returns a set of Recommendation objects. If the recommendations have not yet being computed, the
Each Recommendation refers to the subset recommendations
import event log XStageViewer e g a tse v it c a XLogData
XLogData
XLogData
XLogData XStageManager XLogData
XLogData
XStage
Recommendation
RecommendationFactory active stage generated recommendations RecommendationManager shown in PROCESSEXPLORER which contain the Insight recommendations.
All visualization components, such as the XLogViewer,
ers are separated from the actual recommendation engine. This architecture allows the exploration of different types of visualizations, such as other types of charts, process model visualizations, etc., but keep the actual computation of the recommendations.
In the current implementation of PROCESSEXPLORER, we implemented a multi-perspective trace clustering recommendation engine for subset recommendations and a statistical significance testing approach for obtaining insights recommendations. However, other implementations are easy to implement by extending the corresponding classes.
The PROCESSEXPLORER demo tool can be found at our project page1. On the project page, a demonstration video including a screencast, a reduced event log derived from the BPI Challenge 2019, and additional screenshots are provided. The demo tool requires Oracle Java 8 and was tested on Windows and Ubuntu.
In this paper, we presented PROCESSEXPLORER, an interactive visual recommendation system for process discovery inspired by the workflow typically performed by analysts. Our system suggests two types of recommendations that guide analysts towards interesting subsets of cases as well as shows insightful statistics of relevant PPIs. Subset recommendations are computed using multi-perspective trace clustering to obtain process behavior patterns that are interesting to explore. Insights recommendations show interesting PPIs that significantly differ for an investigated subset compared to the rest of the event log. Furthermore, PROCESSEXPLORER gives each recommendation a score based on interestingness and maturity. It applies top-k diversifying ranking to obtain the most different recommendations.
This work is funded by the German Federal Ministry of Education and Research (BMBF) Software Campus project “AI-PM” [01IS17050] and the research project “KI.RPA” [01IS18022D].