TL;DR
This paper explores how incorporating user and time information into process mining can reveal deeper insights into business process variations. By applying sequence alignment methods to different representations of process data, the authors demonstrate that these additional parameters significantly impact the analysis, leading to a more nuanced understanding of process deviations and user behavior.
What Problem Does This Paper Address?
Process mining aims to discover, monitor, and improve real processes by extracting knowledge from event logs. While existing methods can identify the topological structure of a process, they often overlook crucial contextual information like who performed an activity and how long it took. This limitation can obscure important process variations, user-specific behaviors, and performance bottlenecks. The paper addresses this gap by investigating how to effectively integrate user and time data into sequence alignment techniques to gain a more comprehensive view of process executions.
Methods
The core of this paper lies in applying sequence alignment methods to different "views" of process data. The authors utilize three established sequence alignment techniques:
- Longest Common Substring (LCS): Identifies identical contiguous subsequences. It's sensitive to exact matches and order.
- Longest Common Subsequence (LCSS): Finds the longest subsequence common to both sequences, allowing for intervening elements. This makes it more robust to minor distortions.
- Time-Warped Longest Common Subsequence (TWLCS): Combines LCSS with dynamic time warping to account for non-linear time differences and distortions, making it suitable for comparing sequences with varying durations.
They then define four types of sequences derived from event logs:
- Type A (Events only): Focuses solely on the topological structure of the process.
- Type B (Events + Time): Incorporates the duration of events, represented by repeating event symbols based on predefined duration categories (short, medium, long).
- Type C (Events + Users): Includes information about the users who performed each event.
- Type D (Events + Time + Users): Combines all three aspects for a holistic view.
The paper evaluates how these different sequence types, when analyzed with LCS and LCSS (chosen for their suitability over TWLCS for this specific business process context), reveal distinct patterns and similarities.
Datasets
The study uses a real-world event log from an SAP system detailing an invoice verification process across five European countries. The log spans from January 1, 2012, to June 30, 2012, containing 70,855 records. The data originates from three factories within the company.
Key Insights and Ideas
- Adding context matters: Simply looking at the sequence of events (Type A) misses significant variations. Incorporating time (Type B) and user information (Type C) leads to a much finer-grained analysis and a larger number of distinct case types.
- User behavior is distinct: Analyzing sequences with user information (Type C) reveals how specific users or groups of users contribute to process variations, potentially highlighting knowledge gaps or specialized roles.
- Time reveals process efficiency: Time parameters (Type B) help identify bottlenecks and variations in process execution duration, even for seemingly similar event sequences.
- Combined view is powerful: Type D, which combines topology, time, and user information, offers the most detailed perspective, allowing for the identification of complex scenarios like difficult invoices that take longer to approve.
- Method choice is context-dependent: While TWLCS is powerful, LCS and LCSS proved more suitable for the specific business process data analyzed, suggesting that the choice of sequence alignment method should be tailored to the data and the desired insights.
Key Results
The paper presents results in the form of tables and sequence graphs for each sequence type.
- Type A: Identified CVAAS as the most frequent case type, indicating that most invoices were approved twice. The analysis also highlighted deviations from the "four eyes" rule.
- Type B: Adding time parameters decomposed the most frequent Type A case type into 43 new case types, demonstrating the impact of duration on process variation.
- Type C: Augmenting with user information led to 217 case types. The analysis revealed user patterns, such as specific users frequently initiating and completing processes, and identified a cluster of sequences from a single factory, suggesting potential inter-factory cooperation differences.
- Type D: This combined view further increased the number of distinct case types and revealed interesting findings, such as sequences where a particular user worked on difficult invoices that took longer to approve, even if other users completed similar tasks faster.
While specific quantitative metrics like accuracy improvements are not the primary focus, the qualitative results clearly show that adding user and time parameters significantly increases the granularity and interpretability of process mining analysis.
Limitations and Future Work
The authors acknowledge that the visualization of results for Types B, C, and D was not as clear as for Type A, necessitating a focus on selected case types for detailed analysis. They also note that while the "sequence similarity" metric used for TWLCS doesn't strictly meet all similarity conditions, it was used for simplification.
Future work includes:
- Continuing to explore the impact of different sequence alignment methods and their adjustments on various real-world examples.
- Conducting detailed interpretations of different case studies to determine the optimal method, settings, and views for general data mining and process examination.
- The ultimate goal is to achieve better control over processes by leveraging data from already performed instances.
Summary updated