Text

theses

Process mining techniques attempt to extract non-trivial knowledge and insights from activity logs and use them for further analyses. The traditional process mining focuses on addressing three different problems: workflow discovery, conformance checking and model enhancement. Although many theoretical studies have been done in the process mining domain, studies that applying process mining on solving real-world problems are limited. This dissertation explores how process mining can be used in real-world process analysis to reveal process insights and help human decision making. Novel algorithms and frameworks were proposed to better model and address the real-world problems. In addition, we introduced the recommender system into the process mining domain to help build a data-driven decision support system. Specifically, this dissertation includes three main contributions: (1) application of process mining techniques in real-world medical process analysis; (2) two different process recommender systems; and (3) a process visual analytic tool.
First, we applied process mining techniques to real-world medical process analysis. To enhance the existing workflow discovery algorithm, we developed a splitting-based workflow discovery method. Our method is able to tackle the duplicate-activity problem by allowing the activity nodes in the model to further split. By comparing our discovered model to hand-made expert workflow model of the same process, we were able to find the discrepancies between work-as-done and work-as-imaged. To further quantify and analyze the discrepancies between work-as-done and work-as-imaged, we invented a framework for automatic process deviation detection. Our framework first compares the observed process traces with knowledge-driven workflow models using a phase-based conformance checking algorithm. The discrepancies (process deviations) were analyzed and false alarms were identified. The false alarms were categorized into three types of causes: (1) model gaps or discrepancies between the model (“work as imagined”) and actual practice (“work as done”), (2) errors in activity trace coding, and (3) algorithm limitations. The deviation detection system was then repaired according to the false alarms. With our framework, the deviation detection accuracy was improved from 66.6% to 98.5%. The output system was then applied on unseen datasets to automatically detect the deviations. We applied our framework to two different medical processes and discovered meaningful medical findings. In addition, to analyze the differences between the medical treatment procedures of different patients, we introduced a framework for analyzing the association between treatment procedures and patient cohorts. The framework works by learning weights of context attributes by best-first search, deciding patient cohorts using clustering algorithms, discovering treatment procedures (or patterns) with process mining techniques, and analyzing the cohort-vs.-procedure through statistical analysis.
Second, existing recommender systems have not been developed based on process mining. Our work presents such a bridge. We designed a data-driven process analysis and recommender system that can provide contemporaneous recommendations of process steps and help with retrospective analyses of the process. We first designed a prototype-based recommender system. This approach relies on mining historic data to uncover the potential association between the way of enacting a process and contextual attributes. If association tests are significant, we train a recommender system to output a prototypical enactment for the given context attributes. The system recommends all steps at once. Although it may not be feasible for the performers to study and follow a long list of steps, this recommendation can be used at runtime to automatically verify the process compliance and detect omitted steps and other process errors. Later, we proposed another recommender system that is able to provide step-by-step recommendations. The system was built on recurrent neural networks. The networks took both environmental and behavioral contextual information as input and output next-step suggestions.
Last, we implemented our methods into a visual analytic tool. The tool was named as VIT-PLA, which is short for Visual Interactive Tool for Process Log Analysis. In this tool, we proposed a prototype-based process data visualization strategy. The strategy works by first clustering process data into clusters and then discovering the prototypical procedure from each cluster. Only such cluster prototypes were visualized and presented to the users. Our strategy can greatly reduce the data amount to visualize but preserve the characteristics of each cluster. Statistical analyses were followed and visualized to help analysts better understand their process data.

ETD_9421

Ph.D.

Includes bibliographical references

by Sen Yang

doi:10.7282/t3-arh6-bh20

ETD doctoral

The author owns the copyright to this work.