Pith. sign in

REVIEW 5 major objections 5 minor 48 references

PerFairX: Is There a Balance Between Fairness and Personality in Large Language Model Recommendations?

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Personality-aware prompts in LLM recommenders improve psychological fit but worsen demographic fairness, with DeepSeek and ChatGPT landing on opposite sides of the trade-off.

desk verdict The paper's central trade-off claim is undermined by a circular trait-genre mapping that makes personality alignment largely measure genre overlap, and the empirical base (5 users per dataset) cannot carry the conclusions. read the letter →

arxiv 2509.08829 v1 pith:JZSPJUYB submitted 2025-08-20 cs.CY cs.AIcs.IR

Chandan Kumar Sah This is my paper
classification cs.CYcs.AIcs.IR
keywords LLMrecommendersfairnesspersonalityOCEANmodelpromptsensitivitydemographicparitypersonalizationtrade-offPerFairX
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that injecting a user's inferred Big Five personality traits into an LLM prompt makes the resulting recommendations line up much better with that user's psychological profile, but at the cost of larger disparities across gender, age, and occupation groups. To quantify this, it introduces PerFairX, a benchmark that scores LLM-generated recommendations with ten metrics spanning personality alignment, demographic fairness, intra-list diversity, prompt sensitivity, and accuracy. Applied to ChatGPT and DeepSeek on movie and music data, the framework finds that DeepSeek produces highly personalized, diverse lists that shift strongly with prompt wording, while ChatGPT is more stable and slightly fairer at the group level but less personalized. The broader claim is that 'psychological fairness' is a distinct axis from demographic fairness, and that optimizing one can actively harm the other.

What carries the argument

The load-bearing object is the OCEAN personality vector p_u = [Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism], inferred for each user from genre affinities and rating behavior. PerFairX compares neutral prompts against personality-sensitive prompts, then scores outputs with ten metrics: PAS (cosine similarity between the user's OCEAN vector and the genre vector of recommended items), GPA (genre-personality overlap), DP and EO (group-level fairness), ILF (intra-list diversity), SNSR/SNSV and Jaccard (prompt sensitivity), and Precision/Recall. These are collapsed into a single FPx score. The trade-off claim rests on the contrast between PAS and ILF rising while DP and E

What would settle it

Take a sample of MovieLens or Last.fm users, have them complete a validated Big Five questionnaire, and correlate their self-reported OCEAN scores with the genre-inferred vectors used in this paper. If the correlation is near zero, the PAS numbers and the fairness-personality trade-off are artifacts of the mapping. A simpler check: shuffle the genre-to-trait assignments randomly and rerun the pipeline; if PAS improvements and DP increases persist, the specific mapping is not what drives the result.

Watch

Extended reading notes

Core claim

The central claim is that there is a measurable fairness-personality trade-off in LLM-based recommendation: personality-sensitive prompting substantially raises personality alignment (DeepSeek's PAS jumps from 0.280 to 0.848 on movies), while demographic parity and equal opportunity worsen (DeepSeek's DP rises from 0.237 to 0.726). Across two datasets and both models, sensitive prompts consistently increase within-list diversity but increase between-group disparities. The two models sit at different operating points on this curve: DeepSeek is stronger on alignment, diversity, and accuracy but more prompt-sensitive; ChatGPT is more stable and slightly fairer on group metrics but less personal

Load-bearing premise

The entire comparison rests on assuming that mapping movie and music genres to Big Five traits actually captures a user's personality; if that mapping is wrong or circular, the alignment scores measure the author's labels rather than real psychological fit.

Editorial extensions

If this is right

  • If the trade-off is real, LLM recommenders need explicit fairness-aware prompt design to avoid widening demographic gaps when personalizing.
  • RecLLM benchmarks should include psychological fairness alongside demographic parity; single-score leaderboards can hide the conflict between the two.
  • Model choice changes the operating point: systems wanting stability can choose ChatGPT-like behavior; systems wanting personality fit must accept prompt sensitivity and monitor group bias.
  • PerFairX provides a concrete way to test mitigation strategies: evaluate debiased or fairness-constrained prompts with the same ten metrics.
  • The framework can be extended to other LLMs and domains, letting developers compare how different models navigate the same personalization-fairness frontier.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The genre-to-OCEAN mapping may not capture real user psychology; if the mapping is arbitrary or circular, PAS and the resulting trade-off are artifacts of the label scheme rather than measures of genuine personality fit.
  • Whether sensitive prompting looks 'unfair' depends on which fairness metric you watch: within-list diversity improves while between-group parity worsens, so the same output list can be simultaneously more diverse and more demographically skewed.
  • A testable extension is to separate prompt-wording bias from true personalization: comparing personality-aligned prompts against prompts that simply request trait-consistent genres directly could reveal whether the demographic skew comes from the personality signal or from how the model interprets trait language.
  • The framing of 'psychological fairness' suggests future audits should check not only whether groups get equal exposure, but whether users with different personality profiles receive equally fitting recommendations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper introduces PerFairX, a benchmark framework for evaluating the trade-off between demographic fairness and personality alignment in LLM-based recommender systems. User OCEAN vectors are inferred from genre-affinity and behavioral proxies in MovieLens 10M and Last.fm 360K, then used to construct neutral and personality-sensitive prompts. Recommendations from ChatGPT and DeepSeek are scored with ten metrics (PAS, GPA, DP, EO, ILF, SNSR/SNSV, Jaccard, Precision, Recall), and an FPx aggregator combines them. The central claim is that personality-sensitive prompting improves trait alignment (especially for DeepSeek) but worsens demographic fairness, and that DeepSeek achieves a better overall balance than ChatGPT.

Significance. The question of whether personality-aware personalization conflicts with demographic fairness in LLM recommenders is timely and important. The paper also uses real, public datasets and two deployed LLMs, and the qualitative examples illustrate prompt-induced narrowing. However, the significance hinges entirely on whether PAS and GPA actually measure psychological alignment and whether the reported fairness/accuracy numbers are meaningful. As constructed, the alignment metrics are circular (the same genre-to-OCEAN mapping is used on both user and item sides), the evaluation uses only five users per dataset without error bars, and there is a direct contradiction in the model-comparison text. The empirical evidence therefore does not support the paper's headline trade-off, even though the framework could, in principle, be rebuilt around validated personality instruments and a much larger evaluation cohort.

major comments (5)
  1. [Sec. 3.3-3.4, Eqs. (1)-(2)] PAS and GPA are structurally circular. The user OCEAN vector p_u is inferred from genre-affinity distributions via a hand-assigned genre-to-trait mapping (Sec. 3.3-3.4). The same mapping is then used to construct g_u from the recommended-item genres in Eq. (1), and Eq. (2) uses the same trait-genre sets G_t. Hence PAS and GPA reduce to weighted genre overlap between the user's historical preferences and the recommended genres, relabeled as 'personality.' No external Big Five scores or validation are provided. The observed PAS improvement under sensitive prompting may simply reflect the model returning the user's preferred genres after being told those inferred trait labels. This invalidates the core 'personality alignment' measurements and the trade-off claim.
  2. [Sec. 4, Tables 2-5] The empirical basis is extremely small: only 5 representative user profiles per dataset (10 users total) are used for the LLM evaluation. All tables report point estimates without variance, confidence intervals, or significance tests. The abstract and Sec. 4.1 use phrases like 'significantly improves' and 'consistently outperforms,' but no statistical support is provided. Moreover, it is not explained how group-level DP and EO are computed from such a small sample or which partitions of the sensitive attributes are used. The empirical engine is too fragile to support the paper's quantitative claims.
  3. [Sec. 3.1.3, Eqs. (3)-(4), (9)-(10)] Multiple central metrics are under-specified. EO in Eq. (4) depends on true relevance Y, and Precision/Recall in Eqs. (9)-(10) depend on a relevant-item set Rel, but no definition of relevance is given for LLM-generated top-15 recommendations. Without an operationalization of relevance, the EO, Precision@15, and Recall@15 numbers in Tables 3-4 are not interpretable. Similarly, the computation of DP in Eq. (3) requires a binary recommendation outcome and a sensitive attribute A, but the results do not state which attributes, which group partitions, or how recommendations are binarized. These are load-bearing gaps for the fairness conclusions.
  4. [Sec. 4.3, Table 4 vs. Figure 5] There is a direct contradiction about which model has lower demographic disparity. Table 4 reports DeepSeek with lower DP and EO than ChatGPT in both datasets (e.g., MovieLens DP: 0.726 vs. 0.825; EO: 0.901 vs. 0.952). Yet Sec. 4.3 states that ChatGPT 'exhibits lower bias in some fairness metrics (e.g., DP and EO),' and the Figure 5 caption repeats this claim. This reversal undermines RQ3 and the model-comparison narrative, and it cannot be reconciled from the data as presented.
  5. [Sec. 3.2, Eq. (11), Table 5] The FPx aggregator is not a meaningful scalar score as used. It sums raw metrics with equal weights, but the metrics have very different scales: PAS and DP are bounded in [0,1], ILF is an unbounded entropy, Precision@15/Recall@15 are small fractions, and Jaccard is bounded. No normalization or sensitivity analysis is reported. The resulting PerFairX scores (roughly 2.0-2.9) are dominated by whichever metrics happen to have larger raw magnitudes, so the conclusion that DeepSeek 'provides a better trade-off' is unsupported.
minor comments (5)
  1. [Sec. 3.1.3] SNSR@K and SNSV@K are defined but never reported in the results tables; only Jaccard@15 is used. Either include the results or remove the metrics from the framework description.
  2. [Eq. (5)] ILF is labeled 'Intra-list Fairness' but is just the entropy of the genre distribution. The probability p(g) is not defined (how is it estimated from a top-15 list?), and the entropy is not normalized, making cross-condition comparisons misleading.
  3. [Table 1] The table is corrupted by rendering artifacts (e.g., 'venus-mars', 'birthday-cake', 'globe' symbols) and inconsistent counts (MovieLens '10,000.1k' rating count, Last.fm 598.0k records vs. 360K dataset name). These need cleanup and verification.
  4. [Sec. 3.5] The sensitive prompt examples are not reproducible: the OCEAN vector is not shown, and prompts vary between 'introverted movie lover,' 'agreeable, introverted, emotionally sensitive young female,' and 'soft-spoken, creative young woman.' A precise template that maps p_u to prompt text is needed.
  5. [Fig. 4] The caption says 'Arrows: rightward indicates increased disparity; upward indicates improved alignment,' but the figure's arrows should be individually labeled because not all transitions are rightward (e.g., DeepSeek MovieLens). Consider marking neutral and sensitive endpoints explicitly.

Circularity Check

2 steps flagged · score 6.0 of 10

PAS/GPA are operationalized through the same genre-to-OCEAN mapping used to build the user vector and the prompt, making the 'personality alignment' improvement partly tautological; fairness metrics retain independent content.

  1. renaming known result [Section 3.4 and Eq. (1)]
    "We represent each user u as a five-dimensional vector grounded in the Big Five personality traits (OCEAN model)... we infer these traits using behavioral proxies extracted from user interaction data such as genre affinity distributions. ... P AS(u) = ⃗ pu · ⃗ gu / ∥⃗ pu∥ · ∥⃗ gu∥ (1) where ⃗ pu represents the OCEAN personality vector for user u, and ⃗ gu is the genre vector inferred from the LLM’s recommendation."

    Both the user's OCEAN vector p_u and the recommended items' genre vector g_u are generated from the same hand-assigned genre-to-OCEAN mapping (Sec. 3.3–3.4). p_u is a transform of the user's genre-affinity distribution; g_u is the same transform applied to the recommended genres. PAS is therefore a cosine similarity between two genre-derived profiles, relabeled as 'personality alignment.' No independent Big Five score is measured, and the mapping is not externally validated. The claim that sensitive prompts improve 'psychological fit' reduces to saying that the prompt steers recommendations toward the user's genre history.

  2. self definitional [Section 3.5 and Section 3.1.3, Eq. (1)]
    "Sensitive Prompt Example (for an Introverted User): 'I am an introverted movie lover who prefers thoughtful, emotional stories. Please recommend 15 movies/music.' ... PAS: Measures the cosine similarity between the user’s OCEAN vector ⃗ pu and the inferred genre vector ⃗ gu from the LLM’s recommended items."

    The sensitive prompt is built from the same inferred OCEAN vector that defines the PAS outcome. The prompt injects trait labels derived from the user's genre history, and the score then compares the recommendation's genre-derived trait vector to that same label vector. A high PAS under the sensitive prompt therefore largely reflects the LLM's compliance with the prompt's trait wording, not an independent improvement in psychological alignment. The result is an echo of the input construct: genre → trait → prompt → recommendation → genre → trait → score.

full rationale

The paper's core 'personality alignment' metrics (PAS, GPA) are constructed using a genre-to-OCEAN mapping that is also used to infer the user's personality vector and to formulate the sensitive prompt. This closes the loop: the outcome measure is defined in terms of the same construct that is manipulated, so the reported improvement in alignment is partly a measure of prompt adherence rather than an external validation of psychological fit. The paper itself acknowledges that the mapping is used to 'simulate personality profiles,' which limits the external-validity claim. That said, the fairness metrics (DP, EO, ILF) and prompt-sensitivity metrics are defined independently of the trait mapping, and the observed trade-off between PAS and DP is an empirical property of the LLM outputs, not mathematically forced. The trait-genre mapping is cited to external prior work [14], so the issue is not a self-citation chain. Overall, the central 'personality' result is partially circular by construction, warranting a score of 6 rather than a clean bill of health.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are posited; the 'PerFairX score' is a composite metric, not an entity with an independent falsifiable handle. The ledger captures the hand-assigned mappings, thresholds, and unspecified inference choices that the central claims depend on.

free parameters (5)
  • Genre-to-OCEAN mapping weights
    The mapping from genres to traits in Sec. 3.3 is not given numerically and was not validated; every PAS/GPA value depends on it.
  • OCEAN inference hyperparameters
    Sec. 3.4 says traits are inferred from genre affinity, rating dispersion, temporal activity, and catalog diversity, but the formula and thresholds are unspecified; these choices determine p_u.
  • FPx weights alpha, beta, gamma, delta, epsilon, zeta, eta, mu = all equal (implied by Table 5 note)
    Eq. (11) defines the aggregate with unstated weights; Table 5 says equal weights, i.e., FPx reduces to a raw sum of eight differently-scaled metrics.
  • User interaction cutoff = 200 interactions
    Section 3.3 retains users with at least 200 interactions; this threshold is chosen without analysis and determines which users are in the 5-profile sample.
  • Five representative user profiles per dataset = 5 per dataset
    Section 4 selects 5 profiles per dataset as 'representative' with no stated criteria, no demographic split, and no justification that they can support group-level fairness metrics.
assumptions (5)
  • ad hoc to paper The genre-to-OCEAN trait mapping (e.g., Sci-Fi to Openness, Romance to Agreeableness) is a valid psychometric representation of user personality.
    Introduced in Sec. 3.3-3.4 without validation against any personality questionnaire; it is the basis for both the user profile and the alignment scores.
  • domain assumption A user's Big Five vector can be inferred from behavioral signals (genre affinity, rating dispersion) with no measurement error.
    Sec. 3.4 states the inference method but gives no formula or validation; errors in p_u propagate into every PAS/GPA and prompt.
  • domain assumption The same trait-genre mapping can be applied to LLM-recommended items to obtain a genre vector for cosine alignment.
    Sec. 3.1.3, Eq. (1) computes cosine between p_u and g_u; g_u is inferred from the recommendation list using the same mapping, making alignment a self-consistency check.
  • domain assumption Demographic Parity and Equal Opportunity as defined for classifiers transfer directly to recommendation lists generated for 5 users.
    Eqs. (3)-(4) assume group-level outcome distributions; the paper never explains how these are estimated from a handful of users, yet reports DP/EO in Tables 3-4.
  • domain assumption Relevance labels for Precision@K and Recall@K are available for the LLM's top-15 outputs.
    Equations (9)-(10) require a ground-truth relevance set Rel, but the paper never states how Rel is defined for LLM-generated items.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PerFairX: Is There a Balance Between Fairness and Personality in Large Language Model Recommendations?." pith.science (2026). https://pith.science/paper/JZSPJUYB

@misc{pith2026250908829,
  author       = {Pith},
  title        = {Pith review of: PerFairX: Is There a Balance Between Fairness and Personality in Large Language Model Recommendations?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZSPJUYB}},
  note         = {Machine review of arXiv:2509.08829}
}
read the original abstract

The integration of Large Language Models (LLMs) into recommender systems has enabled zero-shot, personality-based personalization through prompt-based interactions, offering a new paradigm for user-centric recommendations. However, incorporating user personality traits via the OCEAN model highlights a critical tension between achieving psychological alignment and ensuring demographic fairness. To address this, we propose PerFairX, a unified evaluation framework designed to quantify the trade-offs between personalization and demographic equity in LLM-generated recommendations. Using neutral and personality-sensitive prompts across diverse user profiles, we benchmark two state-of-the-art LLMs, ChatGPT and DeepSeek, on movie (MovieLens 10M) and music (Last.fm 360K) datasets. Our results reveal that personality-aware prompting significantly improves alignment with individual traits but can exacerbate fairness disparities across demographic groups. Specifically, DeepSeek achieves stronger psychological fit but exhibits higher sensitivity to prompt variations, while ChatGPT delivers stable yet less personalized outputs. PerFairX provides a principled benchmark to guide the development of LLM-based recommender systems that are both equitable and psychologically informed, contributing to the creation of inclusive, user-centric AI applications in continual learning contexts.

Figures

Figures reproduced from arXiv: 2509.08829 by the authors.

Figure 1
Figure 1. An example of personality-driven movie preferences in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the PerFairX framework. User data from MovieLens and Last.fm are mapped to OCEAN personality vectors, used [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the fairness-personality trade-off in LLM movie recommendations by ChatGPT and DeepSeek. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Trade-off between fairness (DP) and personality align [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of ChatGPT vs. DeepSeek across seven [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 44 canonical work pages

  1. [33]

    Faireval: Evaluating fairness in llm-based rec- ommendations with personality awareness

    Chandan Kumar Sah, Xiaoli Lian, Tony Xu, and Li Zhang. Faireval: Evaluating fairness in llm-based rec- ommendations with personality awareness. arXiv preprint arXiv:2504.07801, 2025. 2, 8

    arXiv 2025

  2. [1]

    A personality-aware group rec- ommendation system based on pairwise preferences

    Roza Abolghasemi, Paal Engelstad, Enrique Herrera- Viedma, and Anis Yazidi. A personality-aware group rec- ommendation system based on pairwise preferences. Infor- mation Sciences, 595:1–17, 2022. 3

    work page 2022

  3. [2]

    Information retrieval meets large lan- guage models: a strategic report from chinese ir community

    Qingyao Ai, Ting Bai, Zhao Cao, Yi Chang, Jiawei Chen, Zhumin Chen, Zhiyong Cheng, Shoubin Dong, Zhicheng Dou, Fuli Feng, et al. Information retrieval meets large lan- guage models: a strategic report from chinese ir community. AI Open, 4:80–90, 2023. 8

    work page 2023

  4. [3]

    User percep- tion of fairness-calibrated recommendations

    Gabrielle Alves, Dietmar Jannach, Rodrigo Ferrari De Souza, and Marcelo Garcia Manzato. User percep- tion of fairness-calibrated recommendations. In Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Personalization, pages 78–88, 2024. 8

    work page 2024

  5. [4]

    Tallrec: An effective and efficient tuning framework to align large language model with recom- mendation

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Tallrec: An effective and efficient tuning framework to align large language model with recom- mendation. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 1007–1014, 2023. 8

    work page 2023

  6. [5]

    Fairness in recommendation ranking through pairwise com- parisons

    Alex Beutel, Jilin Chen, Tulsee Doshi, Hai Qian, Li Wei, Yi Wu, Lukasz Heldt, Zhe Zhao, Lichan Hong, Ed H Chi, et al. Fairness in recommendation ranking through pairwise com- parisons. In Proceedings of the 25th ACM SIGKDD inter- national conference on knowledge discovery & data mining, pages 2212–2220, 2019. 1, 3

    work page 2019

  7. [6]

    Multisided fairness for recommendation

    Robin Burke. Multisided fairness for recommendation. arXiv preprint arXiv:1707.00093, 2017. 1

    arXiv 2017

  8. [7]

    A survey on evaluation of large lan- guage models

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large lan- guage models. ACM transactions on intelligent systems and technology, 15(3):1–45, 2024. 2

    work page 2024

Show all 48 references
  1. [8]

    Bias and unfairness in information re- trieval systems: New challenges in the llm era

    Sunhao Dai, Chen Xu, Shicheng Xu, Liang Pang, Zhenhua Dong, and Jun Xu. Bias and unfairness in information re- trieval systems: New challenges in the llm era. In Proceed- ings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6437–6447, 2024. 2, 4

    2024

  2. [9]

    Fairevalllm

    Yashar Deldjoo. Fairevalllm. a comprehensive framework for benchmarking fairness in large language model recom- mender systems. arXiv e-prints, pages arXiv–2405, 2024. 2

    2024

  3. [10]

    Cfairllm: Consumer fairness evaluation in large-language model recommender system

    Yashar Deldjoo and Tommaso Di Noia. Cfairllm: Consumer fairness evaluation in large-language model recommender system. arXiv preprint arXiv:2403.05668, 2024. 2

    2024 arXiv

  4. [11]

    Fairness in rec- ommender systems: research landscape and future direc- tions

    Yashar Deldjoo, Dietmar Jannach, Alejandro Bellogin, Alessandro Difonzo, and Dario Zanzonelli. Fairness in rec- ommender systems: research landscape and future direc- tions. User Modeling and User-Adapted Interaction, 34(1): 59–108, 2024. 2

    2024

  5. [12]

    Fairness in music recommender systems: A stakeholder-centered mini review

    Karlijn Dinnissen and Christine Bauer. Fairness in music recommender systems: A stakeholder-centered mini review. Frontiers in big Data, 5:913608, 2022

    2022

  6. [13]

    It’s not you, it’s me: the impact of choice models and ranking strategies on gender imbalance in music recommendation

    Andres Ferraro, Michael D Ekstrand, and Christine Bauer. It’s not you, it’s me: the impact of choice models and ranking strategies on gender imbalance in music recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems, pages 884–889, 2024. 2

    2024

  7. [14]

    Personality and taxonomy preferences, and the in- fluence of category choice on the user experience for music streaming services

    Bruce Ferwerda, Emily Yang, Markus Schedl, and Marko Tkalcic. Personality and taxonomy preferences, and the in- fluence of category choice on the user experience for music streaming services. Multimedia tools and applications , 78: 20157–20190, 2019. 5

    2019

  8. [15]

    Bias and fairness in large language models: A survey

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 50 (3):1097–1179, 2024. 2

    2024

  9. [16]

    Intelligent news advertise- ment: A prompt learning approach to personalization using large language models

    Julian G ´omez and Paola Ram´ırez. Intelligent news advertise- ment: A prompt learning approach to personalization using large language models. Eastern European Journal for Mul- tidisciplinary Research, 3(2):325–337, 2024. 2

    2024

  10. [17]

    Personality traits and emotion regulation: A targeted review and recommendations

    David J Hughes, Ioannis K Kratsiotis, Karen Niven, and David Holman. Personality traits and emotion regulation: A targeted review and recommendations. Emotion, 20(1): 63, 2020. 2

    2020

  11. [18]

    Item-side fairness of large language model-based recommendation system

    Meng Jiang, Keqin Bao, Jizhi Zhang, Wenjie Wang, Zhengyi Yang, Fuli Feng, and Xiangnan He. Item-side fairness of large language model-based recommendation system. In Proceedings of the ACM Web Conference 2024, pages 4717– 4726, 2024. 2, 4

    2024

  12. [19]

    Psychology-informed recommender systems tutorial

    Elisabeth Lex and Markus Schedl. Psychology-informed recommender systems tutorial. In Proceedings of the 16th ACM Conference on Recommender Systems, pages 714–717,

  13. [20]

    Psychology-informed recommender systems

    Elisabeth Lex, Dominik Kowald, Paul Seitlinger, Thi Ngoc Trang Tran, Alexander Felfernig, Markus Schedl, et al. Psychology-informed recommender systems. Foundations and trends® in information retrieval, 15(2):134–242, 2021. 3

    2021

  14. [21]

    Improving prompt-based news rec- ommendation with individual template and customized an- swer

    Yijiang Li and Jun Wu. Improving prompt-based news rec- ommendation with individual template and customized an- swer. In Proceedings of the 33rd ACM International Con- ference on Information and Knowledge Management , page 3887–3891, New York, NY , USA, 2024. Association for C...

    2024

  15. [22]

    Towards personalized fairness based on causal notion

    Yunqi Li, Hanxiong Chen, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. Towards personalized fairness based on causal notion. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, pages 1054–1063, 2021. 2

    2021

  16. [23]

    Fairness identifica- tion of large language models in recommendation

    Wei Liu, Baisong Liu, Jiangcheng Qin, Xueyuan Zhang, Weiming Huang, and Yangyang Wang. Fairness identifica- tion of large language models in recommendation. Scientific Reports, 15(1):5516, 2025. 2

    2025

  17. [24]

    Informa- tion retrieval meets large language models

    Zheng Liu, Yujia Zhou, Yutao Zhu, Jianxun Lian, Chaozhuo Li, Zhicheng Dou, Defu Lian, and Jian-Yun Nie. Informa- tion retrieval meets large language models. In Companion 9 Proceedings of the ACM Web Conference 2024, pages 1586– 1589, 2024. 8

    2024

  18. [25]

    Llm-generated expla- nations for recommender systems

    Sebastian Lubos, Thi Ngoc Trang Tran, Alexander Felfernig, Seda Polat Erdeniz, and Viet-Man Le. Llm-generated expla- nations for recommender systems. In Adjunct Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Personalization, pages 276–285, 2024. 8

    2024

  19. [26]

    Llm-rec: Personalized recommenda- tion via prompting large language models

    Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qi- fan Wang, Si Zhang, Ren Chen, Christopher Leung, Jiajie Tang, and Jiebo Luo. Llm-rec: Personalized recommenda- tion via prompting large language models. arXiv preprint arXiv:2307.15780, 2023. 1, 4

    2023 arXiv

  20. [27]

    An introduction to the five-factor model and its applications.Journal of personality, 60(2):175–215, 1992

    Robert R McCrae and Oliver P John. An introduction to the five-factor model and its applications.Journal of personality, 60(2):175–215, 1992. 5

    1992

  21. [28]

    More or less discrim- ination? practical feasibility of fairness auditing of technolo- gies for personnel selection

    Helena Mihaljevi ´c, Ivana M ¨uller, Katja Dill, Aysel Yollu- Tok, and Maximilian von Grafenstein. More or less discrim- ination? practical feasibility of fairness auditing of technolo- gies for personnel selection. AI & SOCIETY , 39(5):2507– 2523, 2024. 8

    2024

  22. [29]

    Data mining: Concepts and tech- niques

    What Is Data Mining. Data mining: Concepts and tech- niques. Morgan Kaufinann, 10(559-569):4, 2006. 4

    2006

  23. [30]

    Fairlens: Auditing black-box clinical decision support systems

    Cecilia Panigutti, Alan Perotti, Andr ´e Panisson, Paolo Ba- jardi, and Dino Pedreschi. Fairlens: Auditing black-box clinical decision support systems. Information Processing & Management, 58(5):102657, 2021. 2

    2021

  24. [31]

    Unveiling bias in fairness evaluations of large language models: A critical literature review of mu- sic and movie recommendation systems

    Chandan Kumar Sah, Lian Xiaoli, and Muhammad Mi- rajul Islam. Unveiling bias in fairness evaluations of large language models: A critical literature review of mu- sic and movie recommendation systems. arXiv preprint arXiv:2401.04057, 2024. 2

    2024 arXiv

  25. [32]

    Navigating the ai frontier: a critical literature review on integrating artificial intelligence into software engineering education

    Chandan Kumar Sah, Lian Xiaoli, Muhammad Mirajul Is- lam, and Md Kamrul Islam. Navigating the ai frontier: a critical literature review on integrating artificial intelligence into software engineering education. In 2024 36th Interna- tional Conference on Software Engineering E...

    2024

  26. [34]

    The next big five inventory (bfi-2): Developing and assessing a hierarchical model with 15 facets to enhance bandwidth, fidelity, and pre- dictive power

    Christopher J Soto and Oliver P John. The next big five inventory (bfi-2): Developing and assessing a hierarchical model with 15 facets to enhance bandwidth, fidelity, and pre- dictive power. Journal of personality and social psychology, 113(1):117, 2017. 2

    2017

  27. [35]

    User modeling in the era of large language models: Current research and future direc- tions

    Zhaoxuan Tan and Meng Jiang. User modeling in the era of large language models: Current research and future direc- tions. arXiv preprint arXiv:2312.11518, 2023. 2

    2023 arXiv

  28. [36]

    Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM conference on recommender systems, pages 269–278, 2020. 2

    2020

  29. [37]

    Fairness matters: A look at llm- generated group recommendations

    Antonela Tommasel. Fairness matters: A look at llm- generated group recommendations. In Proceedings of the 18th ACM Conference on Recommender Systems , pages 993–998, 2024. 2

    2024

  30. [38]

    Counterfactual explanation for fairness in recom- mendation

    Xiangmeng Wang, Qian Li, Dianer Yu, Qing Li, and Guan- dong Xu. Counterfactual explanation for fairness in recom- mendation. ACM Transactions on Information Systems , 42 (4):1–30, 2024. 8

    2024

  31. [39]

    A survey on the fairness of recommender sys- tems

    Yifan Wang, Weizhi Ma, Min Zhang, Yiqun Liu, and Shaop- ing Ma. A survey on the fairness of recommender sys- tems. ACM Transactions on Information Systems , 41(3):1– 43, 2023. 8

    2023

  32. [40]

    Joint multisided exposure fairness for recommen- dation

    Haolun Wu, Bhaskar Mitra, Chen Ma, Fernando Diaz, and Xue Liu. Joint multisided exposure fairness for recommen- dation. In Proceedings of the 45th International ACM SIGIR Conference on research and development in information re- trieval, pages 703–714, 2022. 1, 2, 3

    2022

  33. [41]

    Popularity bias in personal- ity perspective: An analysis of how personality traits expose individuals to the unfair recommendation

    Emre Yalcin and Alper Bilge. Popularity bias in personal- ity perspective: An analysis of how personality traits expose individuals to the unfair recommendation. Concurrency and Computation: Practice and Experience , 35(9):e7647, 2023. 2

    2023

  34. [42]

    Behavior align- ment: a new perspective of evaluating llm-based conversa- tional recommendation systems

    Dayu Yang, Fumian Chen, and Hui Fang. Behavior align- ment: a new perspective of evaluating llm-based conversa- tional recommendation systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and De- velopment in Information Retrieval, pages 2286–2290, 2024. 8

    2024

  35. [43]

    Ensuring user-side fairness in dynamic recommender systems

    Hyunsik Yoo, Zhichen Zeng, Jian Kang, Ruizhong Qiu, David Zhou, Zhining Liu, Fei Wang, Charlie Xu, Eunice Chan, and Hanghang Tong. Ensuring user-side fairness in dynamic recommender systems. In Proceedings of the ACM Web Conference 2024, pages 3667–3678, 2024. 8

    2024

  36. [44]

    Where to go next for recommender systems? id-vs

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. Where to go next for recommender systems? id-vs. modality-based recommender models revisited. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development...

    2023

  37. [45]

    Is chatgpt fair for recommenda- tion? evaluating fairness in large language model recom- mendation

    Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Is chatgpt fair for recommenda- tion? evaluating fairness in large language model recom- mendation. In Proceedings of the 17th ACM Conference on Recommender Systems, pages 993–999, 2023. 1, 2, 4, 8

    2023

  38. [46]

    Towards empathetic conversa- tional recommender systems

    Xiaoyu Zhang, Ruobing Xie, Yougang Lyu, Xin Xin, Pengjie Ren, Mingfei Liang, Bo Zhang, Zhanhui Kang, Maarten de Rijke, and Zhaochun Ren. Towards empathetic conversa- tional recommender systems. In Proceedings of the 18th ACM Conference on Recommender Systems , pages 84–93,

  39. [47]

    Leveraging opposite gen- der interaction ratio as a path towards fairness in online dat- ing recommendations based on user sexual orientation

    Yuying Zhao, Yu Wang, Yi Zhang, Pamela Wisniewski, Charu Aggarwal, and Tyler Derr. Leveraging opposite gen- der interaction ratio as a path towards fairness in online dat- ing recommendations based on user sexual orientation. In Proceedings of the AAAI Conference on Artificial...

    2024

  40. [48]

    Calibrate before use: Improving few-shot perfor- mance of language models

    Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot perfor- mance of language models. In International conference on machine learning, pages 12697–12706. PMLR, 2021. 2 10

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.