Sirvi Autor "Suslova, Viktorija" järgi

Nüüd näidatakse 1 - 1 1

LLMs as a Tool for User Experience Research: A Comparison of Synthetic and Real-World Data
(Tartu Ülikool, 2025) Suslova, Viktorija; Halas, Yana, juhendaja; Eden, Grace, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
This thesis examines whether large language models (LLMs) can effectively support early-stage User Experience (UX) research by generating synthetic user data. The study compares two datasets within the domain of online food delivery services: one from five semi-structured interviews with women aged 35-45 in Tallinn, and another generated using ChatGPT following the same interview guide. Interviews from both datasets were thematically analysed, and the resulting personas and user scenarios were assessed through a side-by-side review focusing on emotional nuance, demographic accuracy, coherence, and the realism of goals, tasks, and interactions. Results show that LLMs can produce coherent, structured, and context-relevant outputs that capture common user concerns such as navigation, delivery timing, and order modifications. Synthetic data proved useful for generating plausible personas and scenarios quickly, offering advantages in speed, consistency, and thematic breadth. However, the outputs often lacked the emotional depth, context-specific details, and behavioural variability present in real-world data. LLMs tended to generalise user behaviour, repeat similar points across participants, and occasionally introduce plausible but unfounded “false positives” that could misdirect research. Persona and scenario comparisons reinforced these trends, showing fewer concrete tasks, omission of some steps, and the inclusion of elements not mentioned by real-world participants. The findings suggest that LLMs can complement, but not replace, traditional qualitative methods. They are best used to accelerate exploration and extend research reach when time, budget, or participant access is limited, provided outputs are refined and validated by human researchers to ensure they reflect genuine user needs.