image 
 
 
     
PhD Dissertation Defense
image.png
Empowering Business Professionals with Interactive What-If Analysis: From Conceptual Framework to LLM-Powered Automation
Sneha Gathani

     
Monday, October 27

11:00:00 AM EST

IRB-4105

https://umd.zoom.us/j/9371793618?pwd=OXhhOXFBUG5uU3dUVndSai9mUWtzdz09&omn=98929182380&jst=2

     
Abstract:

What-if analysis (WIA) is an iterative, multi-step analytical process that enables users to explore and compare hypothetical scenarios by dynamically varying input parameters, applying constraints, and scoping subsets of data to observe their impact on outcomes. It is widely adopted across domains–from business fields like marketing and sales, where organizations evaluate strategies to improve key performance indicators (KPIs), to scientific and engineering fields such as healthcare, climate science, and transportation, where researchers assess alternative hypotheses and interventions. Despite its ubiquity, it remains challenging to conduct WIA effectively: spreadsheet-based tools and business intelligence (BI) platforms require labor-intensive manual configuration, limiting accessibility to experts and hindering scalability across many scenarios; emerging natural language (NL) interfaces powered by AI, while promising broader accessibility, are often brittle, misinterpreting user intent and producing inconsistent results across conversation turns.

This dissertation addresses these challenges by combining human-centered and system-driven approaches, contributing along four interconnected fronts: (1) formalizing a WIA framework, (2) collecting empirical insights from real-world practitioners, (3) developing an interactive system reflecting these insights, and (4) exploring methods for its reliable automation. First, it introduces Praxa, a conceptual framework distilled from a systematic review of 141 publications, which standardizes the vocabulary, fundamental components (what: dataset, features, model; why: objectives; and how: user operations and system operations), and distinct types of WIA. Second, it grounds WIA in practice by examining current tools, workflows, and challenges faced by business professionals (marketing, sales, product, and operations managers), revealing the need for interactive WIA functionalities accessible to non-technical users. Third, guided by these insights, the dissertation presents Decision Studio, an interactive system that makes key WIA functionalities—importance analysis, sensitivity analysis, goal-seek analysis, and constrained analysis—more accessible and effective for reasoning about relationships between drivers (input parameters) and outcomes (KPIs). Finally, it advances automation via a two-stage approach: (1) NL WIA questions are first translated into Praxa Specification Language (PSL)—a declarative, transparent, inspectable, and repairable representation of high-level intent tied to low-level WIA components operationalized from Praxa—using LLMs, and (2) the resulting specifications are compiled into interactive visual interfaces with linked visualizations and parameter controls. Through a benchmark of 405 WIA questions spanning five business datasets and eleven WIA types, the dissertation evaluates three state-of-the-art LLMs, develops a taxonomy of specification errors, proposes error-aware detection methods, and human-guided targeted repair strategies that reduce error rates, demonstrating both the feasibility and improved reliability of LLM-driven WIA automation through human-AI collaboration. By uniting theoretical foundations, empirical insights, interactive system design, and declarative representations and LLM-powered automation, this dissertation lays the groundwork for next-generation interactive analytic systems that democratize complex WIA and empower business professionals to make more informed, data-driven decisions.