Privacy laws now require data minimization, but its broader efects on recommender systems (RS) are still unclear. We systematically study how common minimization techniques reshape the three key RS goals-accuracy, user fairness, and provider fairness. Across multiple datasets and models we (i) measure performance shifts under data minimization strategies, (ii) pinpoint techniques that best balance the three objectives, and (iii) compare model robustness to data reduction. We find that while several strategies improve group-level consumer fairness, they often reduce accuracy and can even worsen provider fairness; the size of these trade-ofs strongly depends on the chosen technique and model. Code and data are public at https://github.com/salvatore-bufi/ DataMinimizationFairness.
Recommender systems (RSs) form the backbone of many digital platforms, tailoring content in
ecommerce, streaming, and social media [
This research carries out a comprehensive investigation into the efects of data minimization on
recommender systems, moving beyond a simple accuracy analysis to consider its broader implications
for both users and item providers. In this section, we detail the experimental setup employed to address
the following research questions:
RQ1: How does data minimization impact the accuracy, provider fairness, and consumer fairness of
recommendations? What are the observable trade-ofs at the strategy level?
RQ2: Which data minimization strategies most efectively preserve recommender system performance
compared to training on the full dataset?
RQ3: How robust are diferent recommendation models to data minimization, particularly concerning
maintaining a balance between accuracy and fairness?
Data Preparation and Protocol. Experiments rely on two public benchmarks: MovieLens 1M (ML1M)
[20, 14] and Ambar[19]. To make data-minimization efects observable while keeping models trainable,
we follow Biega et al. [
Following this pre-processing, we implement our evaluation protocol which adapts the methodology
from Biega et al. [
Algorithms and Metrics. Our study evaluates the impact of data minimization on five representative algorithms from distinct methodological families: the graph-based LightGCN [21], neighborhood-based User-kNN [33], matrix factorization (BPRMF [32]), a linear model (EASER [36]), and a variational autoencoder (MultiVAE [24]). This diverse selection ensures our findings are comprehensive across diferent recommendation approaches. We assess performance across three objectives. We use nDCG [22] to measure accuracy. For fairness, we evaluate consumer fairness with Mean Absolute Deviation (MAD) [11], which quantifies quality disparities between user groups, and provider fairness with Ranking-based Statistical Parity (RSP) [39], which measures how uniformly item categories are exposed. Finally, we use the Hypervolume (HV) [18] to analyze the simultaneous performance on accuracy, consumer, and provider fairness.
Data Minimization Strategies. In our study, we evaluate several approaches to minimize user
data. These strategies are designed to select subsets of user-item interactions that serve as input to the
system. In this paper, we study the data minimization strategies explored by Biega et al. [
This section presents the empirical findings that address our three research questions by analyzing the impact of data minimization on accuracy and fairness2.
Impact of Data Minimization on Key Objectives (RQ1). We find that data minimization significantly impacts recommendation quality, as shown in Table 1. Accuracy consistently drops as fewer interactions are retained (i.e., smaller ), with extreme minimization ( = 1) rendering recommendations impractical. This reduction, however, generally improves consumer fairness (lower MAD) by leveling the performance across user groups, albeit at the cost of relevance. The efect on provider fairness (RSP) is more complex. At low values, the impact is unpredictable, but a clear trade-of with accuracy emerges: strategies that yield higher accuracy, e.g. “Most Rated”, tend to degrade provider fairness by narrowing 1The Most Recent strategy was not applied to the Ambar dataset due to a lack of timestamps. 2Full results on the ML1M and Ambar datasets are available in the original paper.
Strategy while mitigating the fairness degradation seen in other approaches. These intricate trade-ofs are also reflected in the Hypervolume (HV) metric, underscoring that data minimization’s efectiveness cannot be judged on accuracy alone. We conclude that data minimization introduces a critical trade-of: while consumer fairness may improve, it comes at the expense of accuracy and can lead to unpredictable outcomes for provider fairness, requiring nuanced strategies to balance these objectives.
Performance of Data Minimization Strategies (RQ2). To evaluate how well strategies preserve overall performance compared to using the full dataset, we adopt a multi-objective approach. We represent each outcome as a point in a 3D space (nDCG, RSP, MAD) and measure its Euclidean distance ( ) from the reference point (performance on full data). A smaller indicates better preservation. As shown in Figure 1, when is large, most strategies perform similarly well, but a clear hierarchy emerges as decreases. The “Most Rated” and “Most Characteristic” strategies consistently exhibit the highest values, indicating poor overall performance preservation due to their strong biases toward accuracy and fairness, respectively. In contrast, strategies that introduce variability into user profiles, such as “Most Favorite,” and “Highest Variance”, achieve lower values. Thus, data minimization strategies that efectively shape user profiles by introducing variability are best at retaining the holistic performance of the recommender system.
Robustness of Recommender Models (RQ3). To assess model robustness, we analyze the mean ( ) and standard deviation ( ) of the Euclidean distances ( ), across all minimization strategies for a given , as reported in Table 2. A lower signifies better overall robustness, while a lower indicates across all scenarios and proving particularly efective under extreme data sparsity ( = 1) due to its information propagation mechanism. BPRMF also demonstrates strong generalization. In contrast, traditional methods show limitations. UserKNN is robust for moderate minimization ( ≥ falters with scarcer data. EASER, relying on co-occurrence statistics, is highly sensitive to sparsity. 3) but MultiVAE’s performance is dataset-dependent, showing weaker robustness in ML1M but achieving the best performance in Ambar at the global level. We conclude that graph-based modeling, which enhances classic factorization-based representations, is crucial for building recommenders that are robust to extreme data reduction when accuracy and fairness are considered jointly.
We audited the impact of data minimization on the accuracy and fairness trade-ofs in RSs. Our findings reveal a critical trade-of, where consumer fairness often improves at the cost of accuracy and provider fairness. Strategies that introduce variability into user profiles, alongside robust graph-based models, proved most efective at balancing these objectives. Future work will focus on engineering recommender systems that are inherently fair and robust under severe data constraints.
During the preparation of this work, the author did not use any AI tool. systems. In Proceedings of the 33rd ACM Conference on User Modeling, Adaptation and Personalization, UMAP 2025, New York City, NY, USA, June 16-19, 2025, pages 114–123. ACM, 2025. doi: 10.1145/3699682.3728356. URL https://doi.org/10.1145/3699682.3728356. [10] Gemma Galdon Clavell, Mariano Martín Zamorano, Carlos Castillo, Oliver Smith, and Aleksandar Matic. Auditing algorithms: On lessons learned and the risks of data minimization. In AIES, pages 265–271. ACM, 2020. [11] Yashar Deldjoo, Vito Walter Anelli, Hamed Zamani, Alejandro Bellogín, and Tommaso Di Noia. A lfexible framework for evaluating user and item fairness in recommender systems. User Model. User Adapt. Interact., 31(3):457–511, 2021. doi: 10.1007/S11257-020-09285-1. URL https://doi.org/ 10.1007/s11257-020-09285-1. [12] Yashar Deldjoo, Dietmar Jannach, Alejandro Bellogín, Alessandro Difonzo, and Dario Zanzonelli.
Fairness in recommender systems: research landscape and future directions. User Model. User Adapt. Interact., 34(1):59–108, 2024. [13] Dario Di Palma, Giovanni Maria Biancofiore, Vito Walter Anelli, Fedelucio Narducci, and Tommaso Di Noia. Content-based or collaborative? insights from inter-list similarity analysis of chatgpt recommendations. In UMAP (Adjunct Publication), pages 28–33. ACM, 2025. [14] Dario Di Palma, Felice Antonio Merra, Maurizio Sfilio, Vito Walter Anelli, Fedelucio Narducci, and Tommaso Di Noia. Do llms memorize recommendation datasets? A preliminary study on movielens-1m. In SIGIR, pages 2582–2586. ACM, 2025. [15] Tobias Eichinger and Axel Küpper. Distributed data minimization for decentralized collaborative ifltering systems. In ICDCN, pages 140–149. ACM, 2023. [16] Tobias Eichinger and Axel Küpper. On data minimization and anonymity in pervasive mobile-tomobile recommender systems. Pervasive Mob. Comput., 103:101951, 2024. [17] Michèle Finck and Asia Biega. Reviving purpose limitation and data minimisation in personalisation, profiling and decision-making systems. CoRR, abs/2101.06203, 2021. [18] M. Fleischer. The measure of pareto optima. In EMO, volume 2632 of Lecture Notes in Computer
Science, pages 519–533. Springer, 2003. [19] Elizabeth Gómez, David Contreras, Ludovico Boratto, and Maria Salamó. AMBAR: A dataset for assessing multiple beyond-accuracy recommenders. In RecSys, pages 137–147. ACM, 2024. [20] F. Maxwell Harper and Joseph A. Konstan. The movielens datasets: History and context. ACM
Trans. Interact. Intell. Syst., 5(4):19:1–19:19, 2016. [21] Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation. In SIGIR, pages 639–648. ACM, 2020. [22] Walid Krichene and Stefen Rendle. On sampled metrics for item recommendation. In Rajesh Gupta, Yan Liu, Jiliang Tang, and B. Aditya Prakash, editors, KDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, pages 1748–1757.
ACM, 2020. doi: 10.1145/3394486.3403226. URL https://doi.org/10.1145/3394486.3403226. [23] Nian Li, Chen Gao, Depeng Jin, and Qingmin Liao. Disentangled modeling of social homophily and influence for social recommendation. IEEE Trans. Knowl. Data Eng., 35(6):5738–5751, 2023. [24] Dawen Liang, Rahul G. Krishnan, Matthew D. Hofman, and Tony Jebara. Variational autoencoders for collaborative filtering. In WWW, pages 689–698. ACM, 2018. [25] Alberto Carlo Maria Mancino, Antonio Ferrara, Salvatore Bufi, Daniele Malitesta, Tommaso Di Noia, and Eugenio Di Sciascio. Kgtore: Tailored recommendations through knowledge-aware GNN models. In RecSys, pages 576–587. ACM, 2023. [26] Vincenzo Paparella, Vito Walter Anelli, Ludovico Boratto, and Tommaso Di Noia. Reproducibility of multi-objective reinforcement learning recommendation: Interplay between efectiveness and beyond-accuracy perspectives. In Jie Zhang, Li Chen, Shlomo Berkovsky, Min Zhang, Tommaso Di Noia, Justin Basilico, Luiz Pizzato, and Yang Song, editors, Proceedings of the 17th ACM Conference on Recommender Systems, RecSys 2023, Singapore, Singapore, September 18-22, 2023, pages 467–478.
ACM, 2023. doi: 10.1145/3604915.3609493. URL https://doi.org/10.1145/3604915.3609493. [27] Vincenzo Paparella, Vito Walter Anelli, Franco Maria Nardini, Rafaele Perego, and Tommaso Di Noia. Post-hoc selection of pareto-optimal solutions in search and recommendation. In CIKM, pages 2013–2023. ACM, 2023. [28] Vincenzo Paparella, Dario Di Palma, Vito Walter Anelli, and Tommaso Di Noia. Broadening the scope: Evaluating the potential of recommender systems beyond prioritizing accuracy. In Jie Zhang, Li Chen, Shlomo Berkovsky, Min Zhang, Tommaso Di Noia, Justin Basilico, Luiz Pizzato, and Yang Song, editors, Proceedings of the 17th ACM Conference on Recommender Systems, RecSys 2023, Singapore, Singapore, September 18-22, 2023, pages 1139–1145. ACM, 2023. doi: 10.1145/3604915.3610649. URL https://doi.org/10.1145/3604915.3610649. [29] Vincenzo Paparella, Alberto Carlo Maria Mancino, Antonio Ferrara, Claudio Pomo, Vito Walter Anelli, and Tommaso Di Noia. Knowledge graph datasets for recommendation. In Vito Walter Anelli, Pierpaolo Basile, Gerard de Melo, Francesco Maria Donini, Antonio Ferrara, Cataldo Musto, Fedelucio Narducci, Azzurra Ragone, and Markus Zanker, editors, Proceedings of the Fifth Knowledge-aware and Conversational Recommender Systems Workshop co-located with 17th ACM Conference on Recommender Systems (RecSys 2023), Singapore, September 19th, 2023, volume 3560 of CEUR Workshop Proceedings, pages 109–117. CEUR-WS.org, 2023. URL https://ceur-ws.org/ Vol-3560/short2.pdf. [30] Hossein A. Rahmani, Mohammadmehdi Naghiaei, and Yashar Deldjoo. A personalized framework for consumer and producer group fairness optimization in recommender systems. Trans. Recomm.
Syst., 2(3):19:1–19:24, 2024. [31] Theresia Veronika Rampisela, Maria Maistro, Tuukka Ruotsalo, and Christina Lioma. Evaluation measures of individual item fairness for recommender systems: A critical study. CoRR, abs/2311.01013, 2023. [32] Stefen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. BPR: bayesian personalized ranking from implicit feedback. CoRR, abs/1205.2618, 2012. [33] Paul Resnick, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John Riedl. Grouplens:
An open architecture for collaborative filtering of netnews. In CSCW, pages 175–186. ACM, 1994. [34] Abdulaziz Samra, Evgeny Frolov, Alexey Vasilev, Alexander Grigorevskiy, and Anton Vakhrushev.
Cross-domain latent factors sharing via implicit matrix factorization. In RecSys, pages 309–317.
ACM, 2024. [35] Nasim Sonboli, Sipei Li, Mehdi Elahi, and Asia Biega. The trade-of between data minimization and fairness in collaborative filtering. CoRR, abs/2410.07182, 2024. [36] Harald Steck. Embarrassingly shallow autoencoders for sparse data. In WWW, pages 3251–3257.
ACM, 2019. [37] Marco Valentini. Cooperative and competitive llm-based multi-agent systems for recommendation.
In Claudia Hauf, Craig Macdonald, Dietmar Jannach, Gabriella Kazai, Franco Maria Nardini, Fabio Pinelli, Fabrizio Silvestri, and Nicola Tonellotto, editors, Advances in Information Retrieval - 47th European Conference on Information Retrieval, ECIR 2025, Lucca, Italy, April 6-10, 2025, Proceedings, Part V, volume 15576 of Lecture Notes in Computer Science, pages 204–211. Springer, 2025. doi: 10.1007/978-3-031-88720-8\_33. URL https://doi.org/10.1007/978-3-031-88720-8_33. [38] Yifan Wang, Weizhi Ma, Min Zhang, Yiqun Liu, and Shaoping Ma. A survey on the fairness of recommender systems. ACM Trans. Inf. Syst., 41(3):52:1–52:43, 2023. [39] Ziwei Zhu, Jianling Wang, and James Caverlee. Measuring and mitigating item underrecommendation bias in personalized ranking systems. In Jimmy X. Huang, Yi Chang, Xueqi Cheng, Jaap Kamps, Vanessa Murdock, Ji-Rong Wen, and Yiqun Liu, editors, Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, pages 449–458. ACM, 2020. doi: 10.1145/3397271.3401177. URL https://doi.org/10.1145/3397271.3401177.