Use of Air Quality Sensor Network Data for Real-time Pollution-Aware POI Suggestion

Bianca Lau; Eric Morris; Ganga Chinna Rao Devarapu; Giuseppe Fasano; Liam O'Faolain; Sina Adham-Khiabani; Tommaso Di Noia; Xia Liu; Yashar Deldjoo

REVIEW 3 major objections 4 minor 21 references

Use of Air Quality Sensor Network Data for Real-time Pollution-Aware POI Suggestion

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a personalized, privacy-preserving POI recommender can blend user tastes with real-time air-quality scores, and demonstrates a working app that re-ranks nearby places by the tunable combination $S = \alpha…

desk verdict A plausible integration demo whose health-claim is undercut by random AQI values in the only recommendation experiment; worth a workshop-level referee, not more. read the letter →

arxiv 2502.09155 v2 pith:TS2IAHTX submitted 2025-02-13 cs.IR

Giuseppe Fasano , Yashar Deldjoo , Tommaso di Noia , Bianca Lau , Sina Adham-Khiabani , Eric Morris , Xia Liu , Ganga Chinna Rao Devarapu

show 1 more author

Liam O'Faolain

This is my paper

classification cs.IR

keywords recommendersystemspoint-of-interestrecommendationairqualityfederatedlearningpollution-awaremobileappsensornetworkAQI

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

AirSense-R is a mobile recommender that combines two signals for every nearby point of interest: a matrix-factorization score learned from the user's own ratings and a real-time air-quality score drawn from urban sensor networks. The paper's central claim is that a weighted sum of these signals, $S = \alpha S_{\text{MF}} + (1-\alpha) S_{\text{AQI}}$, lets the same app serve both a user who wants the best-rated places and a user who wants the least polluted places, with $\alpha$ as the dial between them. This matters because urban air pollution is uneven at street level, and a destination that looks attractive on a map may sit next to a traffic hotspot. The system also keeps preference data on the device through federated learning, so the health benefit is not bought by sacrificing privacy. The demonstration uses simulated sensor values, but it shows the full pipeline from sensor feed to re-ranked POI list.

What carries the argument

The load-bearing object is the convex combination $S = \alpha S_{\text{MF}} + (1-\alpha) S_{\text{AQI}}$, which re-ranks the candidate points returned by the Places API by merging a preference score and a pollution score under a single user-controlled weight. Around this score sit two supporting mechanisms: a federated-learning scheme in which user embeddings never leave the phone and only shared POI embedding updates are aggregated by Federated Averaging, and an AirSENCE prediction engine that uses FBProphet for trend and seasonality modeling, residual analysis for anomaly detection, and radial-basis-function interpolation to estimate AQI where sensors are sparse.

What would settle it

Recompute the two-user demonstration using recorded AirSENCE AQI values from Bari for the same date and time instead of the random 20 to 70 values, and compare the top-5 lists under $\alpha=0$, $0.3$, $0.5$, and $1$; if the lists barely change or the mean AQI of the recommended POIs does not improve as $\alpha$ moves toward 0, the pollution-awareness claim is not supported.

Watch

Extended reading notes

Core claim

The paper claims that pollution-aware POI suggestion reduces to a tunable re-ranking step. For each candidate POI within a walking radius, the final score is $S = \alpha \cdot S_{\text{MF}} + (1-\alpha) \cdot S_{\text{AQI}}$, where $S_{\text{MF}}$ is the matrix-factorization prediction of user preference and $S_{\text{AQI}}$ is the normalized air-quality index at that location, with missing areas filled by radial-basis-function interpolation. Setting $\alpha=1$ recovers the pure preference ranking, $\alpha=0$ produces a pure air-quality ranking, and intermediate values produce a blend; the demonstration shows that a health-sensitive user with $\alpha=0.3$ receives a different list than a general user, and that training the preference model by federated averaging gives the lowest median absolute error among centralized, distributed, and federated variants. The paper presents this as evidence that real-time sensor data can be turned into personalized, privacy-preserving healthier choices in urban navigation.

Load-bearing premise

The demonstration assumes that eight virtual sensors per user, with AQI values drawn uniformly at random from 20 to 70, behave like real air-quality measurements; if that assumption fails, the observed re-ranking may not occur with actual sensor data.

Editorial extensions

If this is right

With $\alpha$ set to $0$, $0.5$, or $1$, the app produces pure-air-quality, balanced, or pure-preference lists, so health sensitivity can be tuned per user or per health condition.
Because user embeddings stay on-device and only POI embedding updates are shared, the system's privacy guarantee is part of the training loop rather than an external add-on.
Anomaly detection on live sensor streams can flag sudden pollution events, such as the NO spike observed at one Cork device, and push that information into the re-ranking in near real time.
The federated-learning demonstration with held-out users suggests that the privacy-preserving training variant can match or beat centralized training on accuracy, making the approach practical for on-device recommenders.
Interpolating AQI with radial basis functions extends coverage beyond the physical sensor locations, so recommendations remain pollution-aware even where sensors are sparse.

Reading between the lines

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

The paper leaves implicit that the same $\alpha$-weighted score could be applied to walking routes, not just destinations; a path that avoids a polluted street could reduce exposure more than re-ranking the endpoint.
The simulation with uniformly random AQI values cannot establish that real pollution fields change recommendations, because real AQI is spatially correlated and varies with traffic and weather; replaying the demonstration with recorded AirSENCE measurements would be a stronger test.
If the approach scales, the prediction engine's forecasts could turn the system from reactive to proactive, telling a user the least polluted time of day to visit a POI, which the paper names as future work.
The demonstration focuses on two users, so the claimed flexibility across health sensitivity is plausible but not yet shown across a population; a user study with heterogeneous health profiles would test whether the $\alpha$ dial actually changes real-world choices.

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.

Desk Editor's Note

read the letter

What you should know first: this is a demo paper, and the central claim that the system actually delivers pollution-aware POI suggestions is not demonstrated. The only quantitative recommendation experiment uses simulated virtual sensors with AQI values drawn uniformly at random between 20 and 70. With random values, the AQI re-ranking produces an arbitrary ordering, so the intra-user and inter-user demos show only that changing alpha changes the list, not that the lists correspond to lower-pollution choices. The stress-test note is right on this: the real AirSENCE measurements appear only in descriptive plots, not in the recommendation evaluation.

What the paper does well is the integration concept. Merging federated matrix factorization with real-time AQI from a low-cost sensor network for POI re-ranking is a new application relative to the cited literature. The architecture is clean and easy to follow: client-side embeddings, federated averaging, a simple weighted score S = alpha*S_MF + (1-alpha)*S_AQI, and radius-based proximity filtering. The descriptive plots of real sensor data (weekly NO2 diurnal patterns in Bari, the NO spike in Cork, and a FBProphet forecast) are genuine and show the sensor network works. The code and demo video are public, which is a real plus for a demo paper. The authors also explicitly say in the conclusion that further investigation under realistic conditions is needed, which is honest.

The soft spots are exactly where the reader put them. The random AQI simulation is load-bearing and should have been replaced by real sensor traces, or at least a simulation with spatial and temporal correlation. The federated learning comparison has only three users and no statistical tests; the boxplots are illustrative at best. The FBProphet prediction is shown as a plot without forecast error metrics, so we cannot judge its accuracy. The paper is best read as a feasibility demonstration of a plausible architecture, not as evidence that the approach is effective in real conditions. On the other hand, the formula itself is not circular: alpha is an explicit user choice, and the MF and AQI scores are computed independently.

Who gets value from this? People working on context-aware recommender systems, urban health applications, or citizen-science sensor deployments will find the integration useful and may want to borrow the system sketch. Anyone expecting a validated recommendation model will be disappointed. I would not cite it in my own work, but I might bring it to a reading group as an example of a well-scoped demo with a clear gap between claim and evidence.

Recommendation: yes, send it to peer review as a workshop demo. It deserves referee time because the architecture is coherent, the code is available, and the sensor data plots show real deployment. The referee should ask for the random-AQI simulation to be replaced or clearly framed as a mock-up, and should treat the FL comparison as preliminary. Conditional accept is the right outcome, not desk reject.

Referee Report

3 major / 4 minor

Summary. The paper presents AirSense-R (AirTown), a mobile POI recommendation system that combines collaborative filtering with real-time AQI data from AirSENCE sensor networks in Bari and Cork. The proposed architecture includes a client-server design, federated learning for privacy, an FBProphet-based prediction engine, and a scoring formula S = alpha*S_MF + (1 - alpha)*S_AQI for re-ranking nearby points of interest. The demonstration uses two example users, a set of virtual sensors with random AQI values, and a federated learning comparison involving three users. The paper also includes descriptive analyses of real AirSENCE measurements and a forecast plot. The authors conclude that the system supports healthier urban choices while preserving privacy.

Significance. If substantiated, the system would be a useful integration of environmental sensing with personalized recommendation, and the open-source code and demonstration video are valuable assets. The descriptive analyses of real AirSENCE data (Figures 3 and 4) illustrate the potential value of dense sensor networks. However, the central claim that the system delivers pollution-aware recommendations from live sensor data is not established by the current quantitative evaluation, which replaces real AQI values with random values and relies on an extremely small federated learning comparison.

major comments (3)

[Section 3, 'Demonstration and Results'] The recommendation evaluation replaces real AirSENCE measurements with eight virtual sensors per user whose AQI values are drawn uniformly at random from 20 to 70. With random AQI values, S_AQI is uncorrelated with actual pollutant concentrations, so the re-ranking demonstrations in Figure 6 show only how changing alpha reorders lists under noise, not that the lists correspond to lower-pollution choices. Since the abstract and introduction claim that the system merges 'live air quality data from AirSENCE sensor networks' to 'support healthier choices,' this is a load-bearing gap. The evaluation should use actual sensor readings from the Bari/Cork networks, or at least spatially/temporally interpolated values from those sensors, and should report whether the re-ranked recommendations indeed have lower measured or predicted AQI.
[Section 3, 'Federated Learning Demonstration'] The federated learning comparison is based on only three users and reports boxplots of absolute error without any statistical testing, confidence intervals, or per-user error values. The claim that 'the federated model consistently achieves the lowest median absolute error per user' is not supported by a sample size of three, especially without specifying how the 'distributed' scenario differs from the 'federated' scenario. Additional users or repeated subsampling with error bars are needed before drawing any conclusion about the benefit of federated training.
[Section 2, 'AirSENCE Prediction Engine'] The FBProphet prediction is presented only as a plot (Figure 5) with no quantitative forecast metrics such as MAE, RMSE, or predictive intervals, and no evaluation of the claimed anomaly detection or spatial interpolation capabilities. Since the prediction engine is described as a component that enables early detection and provides localized AQI data for recommendations, its accuracy should be reported with concrete numbers, at least on a held-out portion of the AirSENCE time series.

minor comments (4)

[Section 2, Equation (1)] The formula S = alpha*S_MF + (1 - alpha)*S_AQI is never accompanied by definitions of the scales or normalization of S_MF and S_AQI; without normalization the interpretation of alpha as a trade-off weight is unclear.
[Section 3, Figure 6] The figure caption and text refer to 'red-boxed' and 'blue-boxed' lists, but the printed text does not reproduce the figure's colored boxes; please clarify the layout so the description remains meaningful in a monochrome setting.
[Throughout] The system is called AirSense-R in the title and abstract but AirTown in the figures and parts of the text; please use a single consistent name, and fix the typo 'AirTOWNS’s' in Section 4.
[References] Reference [14] contains a malformed URL ('https://http://www.everyaware.eu/...') that should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the recommendation formula is an explicit user-weighted combination of independently computed preference and AQI scores, and the random-AQI evaluation weakness is an evidence limitation, not a circular reduction.

full rationale

The paper's central derivation is the score S = alpha * S_MF + (1-alpha) * S_AQI, where alpha is an explicit user-controlled weight and S_MF and S_AQI are computed from independent sources (matrix factorization over ratings and AQI values). No parameter is fitted to an output and then renamed as a prediction; no uniqueness theorem or load-bearing self-citation is invoked; and no result is defined in terms of the claim it is supposed to support. The evaluation does replace real AirSENCE measurements with random AQI values (Section 3), and the federated-learning comparison appears to include the three target users in the second training step, but these are validity/evidence concerns rather than instances where a derivation reduces to its own inputs. Self-citations appear only in the positioning paragraphs of the conclusion and do not carry the argument. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a user-chosen linear interpolation between two standard models, plus a simulation of the air quality input. The free parameters (alpha, radius, random AQI range) are not fitted to data, but they shape all demonstrated outputs. The main unvalidated assumption is that simulated random AQI values adequately represent real spatiotemporal pollution patterns.

free parameters (3)

alpha (trade-off weight) = 0, 0.3, 0.5, 1 (user-chosen in demo)
User-specified trade-off between preference and air quality; not fitted, but essential to the demonstration's output.
search radius = 1 km
Hard-coded radius restricting candidate POIs; not varied in the demo.
random AQI range for virtual sensors = 20-70
Uniform range used for simulated virtual sensors; arbitrary and not representative of real AQI distributions.

assumptions (5)

domain assumption Air quality affects user health and should influence POI choice.
The entire scoring function rests on this premise; invoked in Section 1.
domain assumption User preferences can be collected locally via surveys and remain on device.
Invoked in Section 2 recommendation engine.
standard math Matrix factorization predicts preferences from rating data.
Standard result, cited to Koren et al. [12].
standard math Federated averaging aggregates client updates to train a shared model.
Cited to McMahan et al. [15].
ad hoc to paper Virtual sensors with random AQI values simulate real sensor coverage for the evaluation.
Introduced in Section 3 and used for all reported results.

invented entities (1)

Virtual sensors
purpose: To generate AQI values around each user in the demo because real sensor coverage is sparse; they are the main input to the recommendation scoring function.
These are simulated data generators, not physical devices; they have no falsifiable handle outside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Use of Air Quality Sensor Network Data for Real-time Pollution-Aware POI Suggestion." pith.science (2026). https://pith.science/paper/TS2IAHTX

@misc{pith2026250209155,
  author       = {Pith},
  title        = {Pith review of: Use of Air Quality Sensor Network Data for Real-time Pollution-Aware POI Suggestion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TS2IAHTX}},
  note         = {Machine review of arXiv:2502.09155}
}

read the original abstract

This demo paper introduces AirSense-R, a privacy-preserving mobile application that delivers real-time, pollution-aware recommendations for urban points of interest (POIs). By merging live air quality data from AirSENCE sensor networks in Bari (Italy) and Cork (Ireland) with user preferences, the system enables health-conscious decision-making. It employs collaborative filtering for personalization, federated learning for privacy, and a prediction engine to detect anomalies and interpolate sparse sensor data. The proposed solution adapts dynamically to urban air quality while safeguarding user privacy. The code and demonstration video are available at https://github.com/AirtownApp/Airtown-Application.git.

Figures

Figures reproduced from arXiv: 2502.09155 by the authors.

**Figure 1.** AirTown integrates real-time data from sensors installed in Bari and Cork, with user preferences to provide personalized, pollution-aware POI recommendations. 1 Introduction Background. Urban air pollution is a major global concern. The World Health Organization (WHO) reports that 99% of the world’s population breathes air that exceeds its recommended limits [1]. This issue is particularly acute in metropolitan are… view at source ↗

**Figure 2.** The architecture of the proposed system. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 4.** NO readings for six AirSENCE devices in Cork, [PITH_FULL_IMAGE:figures/full_fig_p002_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: NO pattern prediction using FBProphet and [PITH_FULL_IMAGE:figures/full_fig_p003_5.png]

**Figure 6.** Figure 6: Demonstration results for AirTown. 4 Conclusion We presented AirSense-R, a mobile application that combines personalization, real-time pollution awareness, privacy, and proximity considerations to deliver health-conscious point-of-interest (POI) recommendations. By in…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 16 canonical work pages

[1]

Air Pollution

Accessed June 16, 2023. Air Pollution. https://www.who.int/health-topics/air- pollution

work page 2023
[2]

Giovanni Maria Biancofiore, Yashar Deldjoo, Tommaso Di Noia, Eugenio Di Sci- ascio, and Fedelucio Narducci. 2024. Interactive Question Answering Systems: Literature Review. Comput. Surveys 56, 9 (2024), 1–38

work page 2024
[3]

Dauge, Mike Kobernus, Arne J

Núria Castell, Hai-Ying Liu, Franck R. Dauge, Mike Kobernus, Arne J. Berre, Josef Noll, Erol Cagatay, and Reidun Gangdal. 2016. Supporting Sustainable Mobility Using Mobile Technologies and Personalized Environmental Information: The Citi- Sense-MOB Approach in Oslo, Norway . Springer International Publishing, Cham, 199–218. https://doi.org/10.1007/978-3-...

work page doi:10.1007/978-3-319-23455-7_11 2016
[4]

Yashar Deldjoo. 2024. Understanding Biases in ChatGPT-based Recommender Systems: Provider Fairness, Temporal stability, and Recency. ACM Transactions on Recommender Systems (2024)

work page 2024
[5]

Yashar Deldjoo and Tommaso Di Noia. 2025. CFaiRLLM: Consumer Fairness Evaluation in Large-Language Model Recommender System. ACM Transactions on Intelligent Systems and Technology (TIST) (2025)

work page 2025
[6]

Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A Review of Modern Recommender Systems using Generative Models (Gen-RecSys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 6448–6458

work page 2024
[7]

Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Ar- nau Ramisa, Rene Vidal, Maheswaran Sathiamoorthy, Atoosa Kasrizadeh, Silvia Milano, et al. 2024. Recommendation with generative models. arXiv preprint arXiv:2409.15173 (2024)

arXiv 2024
[8]

Yashar Deldjoo, Dietmar Jannach, Alejandro Bellogin, Alessandro Difonzo, and Dario Zanzonelli. 2022. A survey of research on fair recommender systems.CoRR (2022)

work page 2022

Show all 21 references

[9]

Yashar Deldjoo, Markus Schedl, Paolo Cremonesi, Gabirella Pasi, et al . 2018. Content-based multimedia recommendation systems: definition and application domains. In Italian Information Retrieval Workshop. 1–4

2018
[10]

European Commission. 2016. Regulation (EU) 2016/679 of the European Parlia- ment and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (Gen...

2016
[11]

Mouzhi Ge, Francesco Ricci, and David Massimo. 2015. Health-aware Food Recommender System. In Proceedings of the 9th ACM Conference on Recommender Systems (Vienna, Austria) (RecSys ’15). Association for Computing Machinery, New York, NY, USA, 333–334. https://doi.org/10.1145/2...

2015
[12]

Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix factorization tech- niques for recommender systems. Computer 42, 8 (2009), 30–37

2009
[13]

Thai, Bryce Christensen, Xiaoting Liu, Matthew Dunbabin, Riki Lamont, Isak Zing, David Wainwright, Christian Witte, Donald Neale, and Lidia Morawska

Tara Kuhn, Rohan Jayaratne, Phong K. Thai, Bryce Christensen, Xiaoting Liu, Matthew Dunbabin, Riki Lamont, Isak Zing, David Wainwright, Christian Witte, Donald Neale, and Lidia Morawska. 2021. Air quality during and after the Commonwealth Games 2018 in Australia: Multiple bene...

2021 doi
[14]

Accessed December 4, 2024

Vittorio Lorenzo. Accessed December 4, 2024. EveryAware: Enhance Environ- mental Awareness through Social Information Technologies. Project Final Report of EU FP7 ICT Project No. 265432. https://http://www.everyaware.eu/resources/ deliverables/D7.3.pdf

2024
[15]

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-efficient learning of deep net- works from decentralized data. In Artificial intelligence and statistics . PMLR, 1273–1282

2017
[16]

Mueller, David Hasenfratz, Olga Saukh, Martin Fierz, and Christoph Hueglin

M.D. Mueller, David Hasenfratz, Olga Saukh, Martin Fierz, and Christoph Hueglin
[17]

Fatemeh Nazary, Yashar Deldjoo, and Tommaso di Noia. 2025. Poison-RAG: Adversarial Data Poisoning Attacks on Retrieval-Augmented Generation in Rec- ommender Systems. arXiv:2501.11759 [cs.IR] https://arxiv.org/abs/2501.11759

2025 arXiv
[18]

M. J. Potosnak, P. Banerjee, M. B. Berkelhammer, R. Sankaran, V. R. Kotamarthi, R. L. Jacob, P. H. Beckman, S. Shahkarami, D. E. Horton, A. Montgomery, and C. E. Catlett. 2019. Array of Things: A high-density, urban deployment of low-cost air quality sensors. In AGU Fall Meeti...

2019
[19]

Dauge, William A

Philipp Schneider, Nuria Castell, Matthias Vogt, Franck R. Dauge, William A. Lahoz, and Alena Bartonova. 2017. Mapping urban air quality in near real-time using observations from low-cost sensors and model information. Environment International 106 (2017), 234–247. https://doi...

2017 doi
[20]

Abu- Mahfouz, Asadullah Shaikh, Mehmet Turan, and Jawad Rasheed

Khawaja Hassan Waseem, Hammad Mushtaq, Fazeel Abid, Adnan M. Abu- Mahfouz, Asadullah Shaikh, Mehmet Turan, and Jawad Rasheed. 2022. Forecast- ing of Air Quality Using an Optimized Recurrent Neural Network. Processes 10, 10 (2022). https://doi.org/10.3390/pr10102117

2022 doi
[2016]

Atmo- spheric Environment 126 (2016), 171–181

Statistical modelling of particle number concentration in Zurich at high spatio-temporal resolution utilizing data from a mobile sensor network. Atmo- spheric Environment 126 (2016), 171–181. https://doi.org/10.1016/j.atmosenv. 2015.11.033

2016 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Air Pollution

Accessed June 16, 2023. Air Pollution. https://www.who.int/health-topics/air- pollution

work page 2023

[2] [2]

Giovanni Maria Biancofiore, Yashar Deldjoo, Tommaso Di Noia, Eugenio Di Sci- ascio, and Fedelucio Narducci. 2024. Interactive Question Answering Systems: Literature Review. Comput. Surveys 56, 9 (2024), 1–38

work page 2024

[3] [3]

Dauge, Mike Kobernus, Arne J

Núria Castell, Hai-Ying Liu, Franck R. Dauge, Mike Kobernus, Arne J. Berre, Josef Noll, Erol Cagatay, and Reidun Gangdal. 2016. Supporting Sustainable Mobility Using Mobile Technologies and Personalized Environmental Information: The Citi- Sense-MOB Approach in Oslo, Norway . Springer International Publishing, Cham, 199–218. https://doi.org/10.1007/978-3-...

work page doi:10.1007/978-3-319-23455-7_11 2016

[4] [4]

Yashar Deldjoo. 2024. Understanding Biases in ChatGPT-based Recommender Systems: Provider Fairness, Temporal stability, and Recency. ACM Transactions on Recommender Systems (2024)

work page 2024

[5] [5]

Yashar Deldjoo and Tommaso Di Noia. 2025. CFaiRLLM: Consumer Fairness Evaluation in Large-Language Model Recommender System. ACM Transactions on Intelligent Systems and Technology (TIST) (2025)

work page 2025

[6] [6]

Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A Review of Modern Recommender Systems using Generative Models (Gen-RecSys). In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 6448–6458

work page 2024

[7] [7]

Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Ar- nau Ramisa, Rene Vidal, Maheswaran Sathiamoorthy, Atoosa Kasrizadeh, Silvia Milano, et al. 2024. Recommendation with generative models. arXiv preprint arXiv:2409.15173 (2024)

arXiv 2024

[8] [8]

Yashar Deldjoo, Dietmar Jannach, Alejandro Bellogin, Alessandro Difonzo, and Dario Zanzonelli. 2022. A survey of research on fair recommender systems.CoRR (2022)

work page 2022