Source-linked AI summary
The Utility of LLMs in Recommender Systems Explanation Evaluation
Kathrin Wardatzky, Oana Inel, Luca Rossetto, Abraham Bernstein
TL;DR
Recommender-system explanation evaluation lacks guidance for choosing methods and reliable, scalable alternatives to exhaustive user studies. This paper generates explanation prototypes and compares evaluations by 14 LLMs with human ratings, finding human-like patterns but low absolute agreement and model- and construct-dependent reliability. The results motivate concise prompts, larger evaluation models, pre-tested constructs, and factuality audits.
Problem
Explanation methods and evaluation options are numerous, but reliable guidance for selecting user-friendly explanations in specific recommender-system settings remains limited.
Method
The study generates 18 explanation prototypes from varied recommender-system and user information, evaluates them with 14 LLMs, and compares ratings with a user study.
Results
Human-like rating patterns and moderate ranking correlations coexist with low overall rating agreement, while larger models align better with human raters.
Takeaways & Limitations
Use concise explanation-generation prompts, prefer evaluation models with at least 120B parameters, pre-test constructs, and audit explanations for factual accuracy.
Takeaways & Limitations
The study uses one explanation-generating model, one dataset, one domain, and path-based explanations, so results may not generalize to other settings.
Abstract
from arXiv · showhide
Explanations play a crucial role in creating trustworthy recommender systems (RS), yet choosing a good explanation method presents challenges. Many explanation methods exist, but little guidance exists on which is best for which setting. Existing explanation generation methods often produce abstract outputs that require further formatting to become user-friendly, with a seemingly endless pool of options. Running user-based evaluations of all possible options is usually unfeasible, while automated evaluation metrics often either assess only the explainer's abstract output or require comparison with a ground truth, which is generally unavailable. Recent studies have shown that large language models (LLMs) can serve as ``judges'' for explanation evaluation, but their reliability has not yet been thoroughly explored. This paper studies the utility of LLMs in selecting an effective explanation method for a given application. We first explore their ability to generate explanation prototypes given varying information about the RS and the user. Specifically, we generate 18 distinct explanation prototypes, which are subsequently evaluated by 14 LLMs of varying sizes across two temperature settings. We compare these against human ratings derived from a user study. Our results show that while LLMs exhibit human-like rating patterns and achieve moderate rank correlation with human raters, their absolute rating agreement is low and varies substantially by model size and evaluation construct. We derive four practical recommendations: keep explanation-generation prompts concise, prefer larger models for evaluation, pre-test evaluation constructs, and audit explanations for factual accuracy, as neither humans nor LLMs reliably detect non-factual content.
1 Introduction
Recommender-system explanation evaluation lacks clear guidance for selecting methods suited to particular applications. This paper investigates whether LLMs can generate and judge explanation prototypes in ways that align with human evaluations.
- Motivation: Explanation evaluation lacks agreed standards, while automated metrics often assess abstract explainer outputs rather than user-facing explanations.This makes it difficult to determine which explanation approach works best for a given application scenario.
- Motivation: LLMs offer a potentially scalable alternative for evaluating explanations because they are comparatively cheap to run at large scale.Prior work has examined their assessment of reasoning, linguistic quality, relevance, persuasiveness, informativeness, transparency, and accuracy.
- Approach: The paper generates 18 user-friendly explanation types by combining four information aspects about the user and recommender system in prompts.The study focuses on explanation prototyping, where many candidate explanations must be narrowed before human evaluation.
- Contribution: The paper examines rating-pattern overlap, ranking, and rating/ranking agreement with human evaluators rather than focusing only on rating accuracy.It also includes a user study comparing human and LLM rating behavior.
- Approach: Fourteen LLMs evaluate the explanations across six quality dimensions, with their ratings compared against human assessments.The contribution spans large and small models and analyzes evaluator reliability.
2 Related Work
Prior work uses LLMs both to generate user-facing recommender explanations and to evaluate them, but explanation assessment still combines imperfect automated metrics with costly user studies. This paper extends LLM-as-a-judge research toward selecting explanations during prototyping.
- LLM-based explanation generation: LLMs have been used to replace explainers or transform explainer outputs into flexible, personalized, and user-friendly explanations.Their effects can depend on application domain and user characteristics, while excessive detail may overwhelm users.
- Explanation evaluation: Automated recommender-system metrics assess properties such as precision, faithfulness, quality, or hallucination, but often target abstract explainer outputs rather than user-facing explanations.Alignment between offline metrics, user perceptions, and explanation goals remains an ongoing research direction.
- Explanation evaluation: User-based evaluation commonly collects perceptions through online studies, A/B tests, controlled experiments, eye-tracking, or card sorting.These approaches provide user-centered evidence but require study participants and experimental resources.
- LLM-as-a-judge evaluation: LLM-as-a-judge studies increasingly compare model scores with ground truth or human annotations and report potentially accurate, cost-efficient evaluation.Prior work also examines aggregation across heterogeneous evaluators.
- LLM-as-a-judge evaluation: LLM judges can exhibit systematic biases, including preferences for outputs from the same model family, later-listed answers, or inconsistent scale usage.These findings motivate examining rating behavior and agreement rather than relying only on scores.
- Research gap: This paper compares user-friendly explanations generated with varied information content against both LLM and human evaluations.It adds analysis of rating behavior and practical guidelines for adopting the LLM-as-a-judge paradigm.
3 Methodology
The study generates path-based movie recommendations and systematically varies prompt information to produce user-facing explanations. Fourteen LLM judges then rate 989 explanations under two temperature settings, alongside a 216-participant user evaluation.
- Recommendation setting: The methodology uses PGPR with TransE embeddings to produce knowledge-graph-based path recommendations and explanations.The selected path-based recommender outputs recommendation paths grounded in structured knowledge.
- Recommendation setting: The Mindreader movie dataset supplies a knowledge graph containing actors, directors, genres, topics, and release decades.The data were split 60/20/20, with recommender and embedding hyperparameters tuned on validation data.
- Recommendation setting: The study generates explanations for the top 10 recommended items and samples 100 test users, yielding 989 user-recommendation pairs after exclusions.Publication years and short movie synopses were added to prompts and user-study instructions.
- Explanation generation: Prompts vary background knowledge, explanation goal, user history, and recommender/path-trace information to examine personalization and goal-specific explanation generation.The design systematically combines these information aspects across explanation conditions.
- Explanation generation: The generation model is gpt-oss 120b at temperature 0, while 14 open-weight and proprietary LLMs evaluate explanations at temperatures 0.0 and 0.3.Temperature 0.3 runs are repeated five times to examine consistency and whether added variance produces more human-like behavior.
- LLM-as-a-judge evaluation: LLM judges rate satisfaction, scrutability, transparency, perceived understanding, persuasiveness, and recommendation quality on five-point Likert scales.The evaluation prompt presents a streaming-service recommendation and asks for ratings and short reasons in JSON format.
- LLM-as-a-judge evaluation: The study deliberately includes conditions lacking truthful user-preference or recommendation information to test whether evaluators detect non-factual explanations.Recommendation-following intention is also measured because persuasive explanations may be non-factual.
- Human evaluation: A user study evaluates selected prompt conditions using a 5 x 4 mixed design, with eight explanations shown per participant.The human study samples five conditions and uses the same six evaluation measures on five-point scales.
4 Results
The results show that prompt information substantially shaped explanation form and ratings, while LLM agreement was generally low and varied across models, temperatures, and constructs. Human ratings also showed substantial disagreement and only modest alignment with LLM rankings.
- Explanation generation: User and RS information changed explanation style, while path-trace prompts produced explanations using path information.Lay-user explanations avoided technical jargon, whereas RS-knowledge prompts used terms such as algorithms, collaborative filtering, and knowledge graphs.
- Explanation generation: Knowl. explanations had the lowest average Flesh Reading Ease score at 16.38, and adding RS information reduced reading ease more than specifying user expertise.Transparency goals did not affect reading ease, while the other prompt elements had statistically significant effects.
- LLM evaluation: 8,944,644 ratings were successfully parsed, representing 99.7% of all model outputs.Unparseable outputs were concentrated in Deepseek and Nemotron because of hallucinations or formatting and rating-scale differences.
- LLM evaluation: LLM rating behavior varied substantially: Claude’s recommendation-quality means ranged from 1.74 to 2.55, while Deepseek’s ranged from 1.33 to 3.11.Temperature 0.3 increased Cogito’s average rating by 0.42, while shifts in other proprietary models remained below 0.1.
- LLM evaluation: LLM agreement was low: Krippendorff’s α never reached 0.67, while large models had the highest mean α at 0.213.The highest agreement concerned perceived recommendation quality, with values ranging from 0.19 to 0.30 across conditions.
- LLM evaluation: User history increased average ratings by 0.25 relative to the path-trace-only baseline, whereas RS information generally lowered ratings.Perceived understanding received the highest mean rating at 4.17; Gemma was most positive and Minimax and Claude most negative overall.
- Human comparison: Human ranking correlations with LLM ratings were significant but small, ranging from Kendall’s τ = 0.077 to 0.240 across conditions.Human raters disagreed more than LLM evaluators in every sampling tier, and humans and LLMs agreed on rank position only for the baseline.
5 Discussion and Recommendations
The paper finds that LLMs can generate and rate recommendation explanations in human-like ways, but alignment depends on prompt design, model size, and evaluation construct. It recommends concise generation prompts, larger evaluators, pre-tested constructs, and factuality audits.
- Explanation generation: Concise prompts are recommended when combining multiple information elements because dense prompts may cause LLMs to ignore user history.The experiments generally supported varying the information used for explanation generation, but combining user history and recommender-system information created this problem.
- LLM rating patterns: Models showed human-like rating patterns, but larger models aligned better with human raters, motivating a recommendation to use models with ≥120B parameters.Models differed in positivity and negativity while still following similar distributional patterns.
- Alignment with humans: Human–LLM agreement was generally low despite moderate ranking correlation, and agreement varied across evaluation constructs.Satisfaction and persuasiveness showed relatively high agreement, whereas scrutability often produced disagreement.
- Alignment with humans: LLM evaluation constructs should be selected and pre-tested because alignment with humans differs by rating dimension.The paper also reports that humans and LLMs preferred plausible explanations lacking factual information over factual alternatives, supporting factuality audits.
6 Limitations and Future Work
The study’s results are limited by its single-model, single-dataset, single-domain design and by the subset of explanation conditions used in the user study. Future work should test broader domains, explanation settings, goals, and evaluation constructs.
- The evaluation uses one model, one dataset, and one domain, so its results might not generalize to other domains or non-path-based explanations.
- The study sampled rating constructs from prior literature, leaving other explanation goals and evaluation constructs for future exploration.
- To keep user-study costs manageable, the researchers evaluated only a subset of explanation conditions, excluding some explanation types.
7 Conclusions
The paper systematically compares LLM- and human-based evaluations of natural-language recommendation explanations. Across generated explanations and six rating dimensions, 14 LLMs showed human-like rating patterns and moderate ranking correlations.
- The study systematically compares LLM- and human-based evaluations of natural-language recommendation explanations.
- Varying information about the recommender system, explanation goal, and user generated explanations from a single explainable recommender system.
- Across six rating dimensions, 14 LLMs showed human-like rating patterns and moderate ranking correlations with human evaluations.