Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PAARS: Persona Aligned Agentic Retail Shoppers

Saab Mansour, Leonardo Perelli, Lorenzo Mainetti, George Davidson, Stefano D'Amato

TL;DR

PAARS proposes a persona-driven framework to simulate retail shoppers with LLM agents, addressing biases and privacy by inducing synthetic personas from anonymized histories. It introduces an alignment suite that evaluates both individual and group similarity to human shoppers, using KL divergence to capture population-level fidelity. Empirical results show that persona conditioning improves query generation, item selection, and session diversity, and demonstrate a preliminary agent-based A/B testing capability. The framework is positioned as scalable and domain-agnostic, with potential applications in offline experimentation, surveying, and inclusivity of underrepresented groups, while highlighting ethical and methodological limitations that warrant careful future work.

Abstract

In e-commerce, behavioral data is collected for decision making which can be costly and slow. Simulation with LLM powered agents is emerging as a promising alternative for representing human population behavior. However, LLMs are known to exhibit certain biases, such as brand bias, review rating bias and limited representation of certain groups in the population, hence they need to be carefully benchmarked and aligned to user behavior. Ultimately, our goal is to synthesise an agent population and verify that it collectively approximates a real sample of humans. To this end, we propose a framework that: (i) creates synthetic shopping agents by automatically mining personas from anonymised historical shopping data, (ii) equips agents with retail-specific tools to synthesise shopping sessions and (iii) introduces a novel alignment suite measuring distributional differences between humans and shopping agents at the group (i.e. population) level rather than the traditional "individual" level. Experimental results demonstrate that using personas improves performance on the alignment suite, though a gap remains to human behaviour. We showcase an initial application of our framework for automated agentic A/B testing and compare the findings to human results. Finally, we discuss applications, limitations and challenges setting the stage for impactful future work.

PAARS: Persona Aligned Agentic Retail Shoppers

TL;DR

PAARS proposes a persona-driven framework to simulate retail shoppers with LLM agents, addressing biases and privacy by inducing synthetic personas from anonymized histories. It introduces an alignment suite that evaluates both individual and group similarity to human shoppers, using KL divergence to capture population-level fidelity. Empirical results show that persona conditioning improves query generation, item selection, and session diversity, and demonstrate a preliminary agent-based A/B testing capability. The framework is positioned as scalable and domain-agnostic, with potential applications in offline experimentation, surveying, and inclusivity of underrepresented groups, while highlighting ethical and methodological limitations that warrant careful future work.

Abstract

In e-commerce, behavioral data is collected for decision making which can be costly and slow. Simulation with LLM powered agents is emerging as a promising alternative for representing human population behavior. However, LLMs are known to exhibit certain biases, such as brand bias, review rating bias and limited representation of certain groups in the population, hence they need to be carefully benchmarked and aligned to user behavior. Ultimately, our goal is to synthesise an agent population and verify that it collectively approximates a real sample of humans. To this end, we propose a framework that: (i) creates synthetic shopping agents by automatically mining personas from anonymised historical shopping data, (ii) equips agents with retail-specific tools to synthesise shopping sessions and (iii) introduces a novel alignment suite measuring distributional differences between humans and shopping agents at the group (i.e. population) level rather than the traditional "individual" level. Experimental results demonstrate that using personas improves performance on the alignment suite, though a gap remains to human behaviour. We showcase an initial application of our framework for automated agentic A/B testing and compare the findings to human results. Finally, we discuss applications, limitations and challenges setting the stage for impactful future work.

Paper Structure

This paper contains 30 sections, 7 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Persona Mining
  4. Alignment Evaluation Suite
  5. Individual and Group alignment
  6. Computing KL divergence
  7. Task definitions
  8. Query generation
  9. Item selection
  10. Session generation
  11. Experimental Results
  12. Query generation
  13. Item selection
  14. Session generation
  15. A/B testing simulation
  16. ...and 15 more sections

Figures (3)

  • Figure 1: The PAARS framework: we synthesize personas from anonymised human shoppers sessions, generate shopping sessions by powering LLM based agents with personas and retail tools, and measure individual and group alignments to ensure reliability of the persona powered agents. Our framework sets the stage for further persona and session generation improvements, with impactful applications in retail and other domains such as agentic A/B testing and surveying tools.
  • Figure 2: Query generation task: we compare agents with and without personas, by measuring the cosine similarity of the agentic queries against the human ones across different query perplexity levels.
  • Figure 3: Search rank distribution of viewed items comparing human behavior to agents with/without personas.