Abstract
Process mining often involves navigating complex event logs to find meaningful patterns. This paper introduces tool support for Cohort Identification, an automated technique that identifies and recommends groups of cases (cohorts) exhibiting behavior significantly different from the general population. By integrating this method into ProM, visual Miner, and Course Insights, the authors provide analysts with actionable insights into process deviations without the need for exhaustive manual filtering.
The Challenge of Complex Event Logs
In process mining, event logs are used to understand real-world business processes, such as insurance claims or student enrollments. However, these processes are often messy and complex. Analysts typically attempt to simplify their analysis by "drilling down"—filtering data by specific attributes like customer type or geographic region.
The primary challenge lies in selection: with hundreds of possible attribute combinations, manually checking every subset to see if "Gold Customers in the West" behave differently than "Silver Customers in the East" is both tedious and prone to error. This research addresses the need for automated guidance to identify the most "interesting" and behaviorally distinct subsets of data.
Automated Cohort Identification Methodology
The authors implement a technique called Cohort Identification, which operates through three main steps:
- Defining Cohorts: The system uses trace attributes (e.g., "claim amount," "gender," "major") to split the log into a "cohort" (cases possessing the attribute) and an "anti-cohort" (all other cases).
- Measuring Distance: A distance measure between 0 and 1 is calculated. A score of 1 indicates the two groups follow entirely different paths, while 0 suggests their behavior is indistinguishable from a random split.
- Ranking: The system provides a ranked list of attribute-value pairs, allowing analysts to focus on groups with the highest behavioral variance.
Tool Support and Implementation
The cohort identification method has been integrated into three distinct environments to cater to different user needs:
- ProM: A stand-alone plug-in designed for researchers and power users, providing a detailed table of ranked cohorts.
- visual Miner (vM): An interactive tool that allows users to click a recommended cohort to instantly filter and update the visual process map. Cohort analysis interface in visual Miner showing ranked behavioral deviations. Source, p. 3
- Course Insights: A learning analytics dashboard designed for educators to identify student groups (e.g., by enrollment status or program) who interact with course materials in unique ways. Course Insights dashboard recommending cohorts based on student attributes. Source, p. 4
Key Insights and Results
The integration of automated cohort identification offers several advantages:
- Automated "Drill-Down": Instead of analysts manually testing hypotheses, the tool proactively identifies where significant behavioral differences exist.
- Explainability: Results are provided in plain English (e.g., "paymentAmount < 36.0"), making insights immediately actionable for business users or teachers, unlike black-box clustering algorithms.
- Quantifiable Differences: The distance metric provides an objective ranking. In practice, cohorts showing a distance of ~0.69 indicate strong behavioral deviations that warrant immediate investigation.
- Cross-Domain Utility: The research demonstrates that the same underlying algorithm is effective for both industrial process optimization and educational learning analytics, as evidenced by a study involving 875 diverse students in Course Insights.
Limitations and Future Work
While the current implementation is robust, the authors identify several areas for future development. For the Course Insights tool, work continues on making the dashboard "fully process aware" to help instructors better visualize specific student journeys. Additionally, the authors plan to develop techniques that explain how cohorts differ (e.g., identifying specific steps skipped or repeated). Finally, while the current implementation is multithreaded, scaling the exhaustive search of attribute combinations for massive logs remains a computational challenge.
Summary updated