2025 Theses Doctoral

Something Similar: Developing Human-Centered Recommendations to Support "in-the-moment" Meal Choices for Health Goal Attainment

Desai, Pooja M.

Chronic disease remains one of the most pressing health challenges today, and nutrition plays a central role in its management. Yet, effective nutrition management is difficult for individuals to sustain. Existing informatics tools tend to focus on planning and tracking, but provide little support for real-time decisions — the very moments when people are most likely to abandon their goals. This dissertation addressed that gap by exploring similarity-based recommendations to help people make meal choices that align with their health goals while still fitting into their existing plans and contexts, in the moment when decisions are made.

Aim 1 investigated how people reason about meal similarity and substitutability, and whether computational methods could reflect that reasoning. Human participants rated meal pairs and created a gold-standard dataset. Findings showed that similarity judgments drew on multiple dimensions, including form, cuisine, macronutrients, and ingredients, and that meals judged as more similar were also considered more substitutable, though context shaped viability. Computational methods, especially LLM-based approaches, closely aligned with human judgments, suggesting that they can generate interpretable, human-aligned similarity assessments.

Aim 2 characterized user needs and assessed the feasibility of similarity-based recommendations in everyday contexts. Findings showed that meal choices were shaped by both personal priorities (taste, cost, preparation time) and situational setting. In open settings, people balanced multiple priorities with the help of existing tools. However, in constrained settings, they made on-the-spot compromises and found similarity-based recommendations particularly useful. Participants also emphasized the importance of autonomy and control at the system, algorithm, and moment-to-moment levels, highlighting the need for flexible, context-aware systems that adapt to evolving goals and constraints.

Finally, Aim 3 explored how large language models (LLMs) could enrich meal data and support conversational recommendations. Study 1 showed that LLMs could generate structured tags from free-text data, creating an enriched corpus that supported tailored recommendations. Study 2 used participatory design to identify preferences for short, user-led dialogues that allowed refinement along attributes such as flavor, texture, macronutrients, and effort while maintaining a supportive tone. Study 3 tested two conversational prototypes and found that participants qualitatively preferred the Dimension-Elicitation Chatbot, valuing its structured, collaborative style. At the same time, technical breakdowns underscored the fragility of current systems and the need for more adaptive and generative approaches.

Together, these studies contribute a new approach to meal recommendation that is grounded in human reasoning, demonstrates the feasibility of using free-text data as both input and corpus, and highlights the unique value of similarity-based methods for in-the-moment nutrition support. More broadly, this dissertation advances human-centered AI by showing how technical methods, human priorities, and ethical safeguards can come together to create decision-support systems that are interpretable, actionable, and designed for real-world use.