Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From latent factors to language: a user study on LLM-generated explanations for an inherently interpretable matrix-based recommender system

Maxime Manderlier, Fabian Lecron, Olivier Vu Thanh, Nicolas Gillis

TL;DR

The paper tackles the challenge of making inherently interpretable matrix-based recommendations understandable to users by employing carefully designed prompts to generate natural-language explanations from large language models. It uses bounded simplex-structured matrix factorization (BSSMF) to obtain interpretable latent user types and then translates these signals into explanations, evaluating four explanation strategies through a large user study. Key findings show that while interpretable models provide strong relevance signals even without explanations, explanations are generally well received but yield only modest differences across strategies; faithful, history-based, and hybrid explanations each have distinct strengths and limitations. The study offers practical guidance on designing faithful and effective explanations, highlights the importance of diversity and personalization, and suggests avenues for adapting explanations to different domains. Overall, the work advances explainable AI by integrating mathematical transparency with user-centered, language-based justification, while providing data, prompts, and analysis code for reproducibility.

Abstract

We investigate whether large language models (LLMs) can generate effective, user-facing explanations from a mathematically interpretable recommendation model. The model is based on constrained matrix factorization, where user types are explicitly represented and predicted item scores share the same scale as observed ratings, making the model's internal representations and predicted scores directly interpretable. This structure is translated into natural language explanations using carefully designed LLM prompts. Many works in explainable AI rely on automatic evaluation metrics, which often fail to capture users' actual needs and perceptions. In contrast, we adopt a user-centered approach: we conduct a study with 326 participants who assessed the quality of the explanations across five key dimensions-transparency, effectiveness, persuasion, trust, and satisfaction-as well as the recommendations themselves. To evaluate how different explanation strategies are perceived, we generate multiple explanation types from the same underlying model, varying the input information provided to the LLM. Our analysis reveals that all explanation types are generally well received, with moderate statistical differences between strategies. User comments further underscore how participants react to each type of explanation, offering complementary insights beyond the quantitative results.

From latent factors to language: a user study on LLM-generated explanations for an inherently interpretable matrix-based recommender system

TL;DR

The paper tackles the challenge of making inherently interpretable matrix-based recommendations understandable to users by employing carefully designed prompts to generate natural-language explanations from large language models. It uses bounded simplex-structured matrix factorization (BSSMF) to obtain interpretable latent user types and then translates these signals into explanations, evaluating four explanation strategies through a large user study. Key findings show that while interpretable models provide strong relevance signals even without explanations, explanations are generally well received but yield only modest differences across strategies; faithful, history-based, and hybrid explanations each have distinct strengths and limitations. The study offers practical guidance on designing faithful and effective explanations, highlights the importance of diversity and personalization, and suggests avenues for adapting explanations to different domains. Overall, the work advances explainable AI by integrating mathematical transparency with user-centered, language-based justification, while providing data, prompts, and analysis code for reproducibility.

Abstract

We investigate whether large language models (LLMs) can generate effective, user-facing explanations from a mathematically interpretable recommendation model. The model is based on constrained matrix factorization, where user types are explicitly represented and predicted item scores share the same scale as observed ratings, making the model's internal representations and predicted scores directly interpretable. This structure is translated into natural language explanations using carefully designed LLM prompts. Many works in explainable AI rely on automatic evaluation metrics, which often fail to capture users' actual needs and perceptions. In contrast, we adopt a user-centered approach: we conduct a study with 326 participants who assessed the quality of the explanations across five key dimensions-transparency, effectiveness, persuasion, trust, and satisfaction-as well as the recommendations themselves. To evaluate how different explanation strategies are perceived, we generate multiple explanation types from the same underlying model, varying the input information provided to the LLM. Our analysis reveals that all explanation types are generally well received, with moderate statistical differences between strategies. User comments further underscore how participants react to each type of explanation, offering complementary insights beyond the quantitative results.

Paper Structure

This paper contains 27 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Recommender system
  5. From ratings to recommendations
  6. Generating explanations
  7. Generating explanations in practice
  8. Explaining the user types
  9. Explaining the recommendations
  10. Study Design
  11. Selecting a representative and usable movie subset
  12. Selecting a consistent user base
  13. Collecting data
  14. Collecting ratings to train the recommendation algorithm
  15. Evaluating recommendations and explanations
  16. ...and 12 more sections

Figures (3)

  • Figure 1: Comparison of the number of times each movie is recommended, with and without the sampling strategy.
  • Figure 2: Mean questions scores across the four experimental conditions.
  • Figure 3: Cliff’s delta for all significantly different pairs identified by Dunn’s test. Values indicate both the direction and magnitude of the effect between groups.