Improving Web Automation Tools through UI Context and Demonstration

Abstract

User interface (UI) automation allows people to perform UI tasks programmatically and can be helpful for computer or smartphone tasks that are tedious, repetitive, or inaccessible. UI automation works by programmatically mimicking a user's interactions on a UI, for example clicking a button or typing into a text field. Traditionally people create UI automation macros by writing code, which requires programming expertise and familiarity with UI technologies. Researchers have explored direct manipulation interfaces and programming-by-demonstration (PBD) to make creating UI automation more accessible for people with less programming experience. With PBD, the user provides demonstrations of how they want their program to behave in a small set of scenarios, and the system then infers a generalized program. Since demonstrations are inherently ambiguous, a key challenge of PBD is in correctly inferring the user's intent and effectively communicating those inferences back to the user. In this thesis, I address important challenges in authoring UI automation macros by leveraging user-provided demonstrations and parameters, and structural patterns in the UI to infer generalized automation; and in understanding UI automation macros by (a) highlighting selected elements on the target UI, (b) visualizing high-level behavior through sequences of actions and UIs visited, (c) visualizing generalizations through color-coding UI elements and grouping corresponding UIs, and (d) providing feedback on validity and uniqueness of element selection logic. First, I conducted two studies observing how programmers write automation code. One of the key challenges participants experienced was in identifying appropriate UI element selection logic. Next, I designed two programming-by-demonstration systems, ParamMacros and ScrapeViz, that enable users to create automation macros without writing code. Users provide demonstrations of what UI elements they want to click or scrape, and then these systems leverage structural patterns in the website DOM to identify patterns and infer generalized automation. ParamMacros supports parameterized macros (powered by user-provided parameters) while ScrapeViz supports distributed hierarchical web scraping macros. ScrapeViz also provides visual tools to help users understand automation behavior in the context of the page source and across different UI pages. This thesis contributes learnings about the challenges users face in creating UI automation macros, and no-code authoring tools and visual understanding tools which have the promise to make UI automation more accessible to a wider audience.