Robotic Process Mining: Accelerating the Adoption of Robotic Process Automation

Completed under a Cotutelle arrangement between the University of Melbourne and University of Tartu© 2021 Volodymyr LenoNowadays, a vast majority of business processes involve many tedious and repetitive tasks, for example transferring data across multiple systems or transforming data from one format to another. Automating these tasks can bring many benefits such as error-rate reduction, performance improvements, workflow standardization and transparency. Automating repetitive tasks has been made more convenient by recent technological advances, particularly the emergence of Robotic Process Automation. Robotic Process Automation (or RPA for short) is a technology that allows organizations to automate routine digital tasks by executing software scripts that encode sequences of fine-grained interactions with Web and desktop applications. Although RPA tools allow automating a wide range of routines, identifying and scoping routines that can be automated is time-consuming. Manual identification of candidate routines via interviews, walkthroughs, or job shadowing allows analysts to identify only the most visible routines. These methods are not suitable for identifying routines for automation in large-scale settings. The thesis studies the problem of discovering routines that can be automated using RPA tools from logs of user interactions with IT systems. This problem is decomposed into a sequence of subproblems such as logs collection and pre-processing, identifying candidate routines, evaluating their automatability, and synthesizing executable specifications that implement them. The thesis provides an overview and analysis of the state-of-the-art for each identified subproblem and identifies gaps to be addressed. Based on this analysis, the thesis makes four contributions. The first contribution is a format to store user interactions logs and a tool to record them. The analysis of existing solutions shows that they are not capable of recording user interactions logs that can be used to discover useful routines. Specifically, they record user interactions at a low level of granularity (keystrokes and clickstreams) and do not capture the data values used during such interactions. The format proposed in the thesis specifies what information is required to be captured by a logging tool to analyze, improve and automate the underlying user interactions. The devised recording tool records user interactions performed in Chrome and MS Excel applications. It records user interactions at a level of logical input elements of a target application. Moreover, it captures all the data values used during these interactions so that the corresponding routines can be analyzed with respect to their automatability and implemented in the form of executable scripts. The second contribution is an end-to-end pipeline to automate routine tasks. Given a user interactions log, the proposed pipeline aims at identifying automatable routines and their boundaries, collect variants of each identified routine, standardize and streamline the identified variants, and discover an executable specification corresponding to a streamlined and standardized variant of the routine. The routines produced as output are defined in a platform-independent language that can be compiled into a script and executed in a target RPA tool. The final two contributions instantiate the proposed pipeline. Specifically, the devised approaches focus on discovering routines for RPA. The first approach aims at discovering candidate routines for RPA from user interactions logs. It applies graph theory to split a user interaction log given as input into a set of segments, each representing an instance of task execution, and then identifies candidate routines by mining frequent sequential patterns from these segments. The second approach focuses on determining whether candidate routines are automatable and discovering their executable specifications. This approach exploits state-of-the-art data transformation discovery techniques and introduces several optimizations to improve their performance and discovery quality. The thesis also describes a method for detecting and removing semantically equivalent routines that lead to the same final effects. Altogether, these approaches give rise to a new family of techniques, called Robotic Process Mining (RPM). RPM aims at helping RPA developers and analysts in the early stages of the RPA lifecycle. Specifically, it assists the analysts in drawing a systematic inventory of the tasks that can be automated with RPA and synthesizes executable specifications of such tasks that can be used as a starting point for their automation. All the artifacts designed and developed for this thesis are publicly available as standalone open-source command-line applications. They are also combined into the pipeline in the form of a software-as-a-service tool called Robidium. The efficiency and potential usefulness of the proposed techniques have been evaluated using synthetic and real-life user interactions logs. The evaluation results show that these approaches can discover automatable routines that present in a user interaction log and identify automatable routines that users recognize as such in real-life logs