{"id":"4df96daa-e17e-4747-b416-c602e7794517","arxiv_id":"2501.13608","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"AirTOWN is a mobile recommender that re-ranks point-of-interest suggestions using live air quality, personal preferences, and distance, with federated learning for privacy.","lead":"This paper describes AirTOWN, a mobile app that recommends restaurants and other places while factoring in real-time air quality around each spot. It combines standard recommendation technology with pollution sensor data and a privacy-friendly training method so users can avoid unhealthy areas without sharing their personal preferences.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core re-ranking S = alpha*SMF + (1-alpha)*SAQI lacks any stated normalization of its components; unless SMF and AQI are on comparable scales, the claimed preference/pollution balance is a scale artifact.","rationale":"Reading the paper as a demo/system description, the central claim is that AirTOWN balances personalization, pollution levels, and distance in its recommendations. The formula S = alpha*SMF + (1-alpha)*SAQI is the only mechanism presented for that balance, so the properties of this formula are load-bearing. The paper does not state the scale of SMF, does not specify any normalization of SAQI or SMF, and does not report numeric score values in the demo. Without scale alignment, alpha is not an interpretable tradeoff parameter. This is a more fundamental issue than the synthetic AQI and two-user limitation the reader flagged: even with real sensor data and a larger user study, the evaluation would not establish balancing unless the components are comparable. The demonstration in Section 3.1, which shows different lists for alpha=0, 0.5, and 1, is consistent with a scale artifact rather than a genuine compromise. The anonymous repository may contain the missing normalization, but the paper itself does not say so, making the central mechanism underspecified. I do not see this as a reason to reject outright, because a short demo paper can be repaired by adding the normalization detail and a quantitative sensitivity check. The privacy claim is also underspecified, but the paper explicitly defers differential privacy to future work, and the re-ranking formula is more central to the demonstrated functionality. Therefore the reader's CONDITIONAL verdict remains appropriate, with the condition tightened to require explicit normalization of the re-ranking components and a quantitative check that alpha actually controls the tradeoff. My agreement with the reader is partial: we both locate the weakness in the evidence for effectiveness, but I would prioritize the internal scale mismatch over the synthetic-data issue as the primary correctness risk.","tokens_in":3789,"tokens_out":4080,"duration_ms":38524,"concrete_test":"Inspect the linked repository (anonymous.4open.science/r/Airtown-Application) for the code that computes S; check whether SMF and SAQI are z-score/min-max normalized or otherwise calibrated before summation. Then re-run the Section 3 demo on the same synthetic data with unnormalized components, using a typical MF rating range (e.g., 1-5) and AQI 20-70, and compute rank correlation between the alpha=0.5 list and each extreme list. If the alpha=0.5 list is near-identical to alpha=0 or alpha=1, the claimed balancing is not demonstrated. The paper should either report this normalization explicitly or replace the formula with a calibrated combination (e.g., percentile ranks or softmax).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 defines the recommendation score as S = alpha*SMF + (1-alpha)*SAQI and describes alpha as modulating the strength of AQI influence. No normalization, standardization, or calibration of the two components is specified. MF rating predictions and AQI readings are not naturally commensurate: ratings often live on a 1-5 scale while AQI values in the demo range from 20 to 70 (and can reach hundreds in real cities). With a naive weighted sum, the component with the larger numeric range will dominate for any fixed alpha, so alpha=0.5 does not mean 'equal influence' unless both terms are brought to a common scale. The Section 3 demonstration reports only qualitative list screenshots for alpha in {0, 0.5, 1}; it never reports the actual S values, the ranges of SMF and SAQI, or a normalization step. This matters because the paper's central claim is precisely that AirTOWN balances user preferences and air quality. If the implementation normalizes internally, the paper needs to say so; if it does not, the 'balanced' list at alpha=0.5 is an artifact of whichever term has larger variance, not evidence of a tunable tradeoff. This is an internal underspecification of the core formula, distinct from (though compounding) the reader's concern about synthetic AQI and two simulated users: even with real data and many users, the re-ranking claim cannot be evaluated without scale alignment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This demo paper presents AirTOWN, a mobile POI recommender that combines matrix-factorization collaborative filtering with real-time Air Quality Index (AQI) readings, using federated learning to keep user embeddings on-device. The system re-ranks candidate POIs within a 1 km radius using the weighted score S = alpha*SMF + (1-alpha)*SAQI, where alpha modulates the influence of AQI relative to user preference. Radial basis function interpolation is used to estimate AQI in areas with sparse sensor coverage. The evaluation is an intra-user and inter-user demonstration in Bari with two simulated users and synthetic AQI values randomly assigned in the range 20 to 70 on a 1x1 km virtual grid; the reported results are qualitative screen captures showing that different alpha values and different user sensitivity settings change the recommendation lists. The paper concludes that initial experiments highlight effectiveness, and it defers differential privacy and larger user studies to future work.","tokens_in":4085,"tokens_out":4788,"duration_ms":43039,"significance":"The integration of real-time environmental data with personalized POI recommendation is timely and relevant, and the architectural choices—matrix factorization for on-device tractability, federated averaging with only item-embedding sharing, and RBF interpolation for sparse sensors—are clearly described and plausible for a mobile demonstration. If the re-ranking were properly calibrated and evaluated, AirTOWN would be a useful demonstrator of health-conscious, privacy-aware navigation. However, the paper provides no quantitative evaluation, no baselines, no privacy analysis, and no reproducible experimental protocol beyond a short qualitative scenario. The core re-ranking formula's scale alignment is underspecified, and the effectiveness and privacy claims in the title, abstract, and conclusion are therefore not currently substantiated. The contribution is best characterized as a system architecture and a proof-of-concept scenario rather than an established effectiveness or privacy result.","major_comments":[{"comment":"The formula S = alpha*SMF + (1-alpha)*SAQI is presented without any normalization, standardization, or calibration of its two components. Matrix-factorization preference scores and AQI readings are not naturally commensurable: user-preference scores often live on a 1-5 scale, while the demonstration uses AQI values in [20, 70] and real-world AQI can reach several hundred. Under a naive weighted sum, the component with the larger numeric range or variance will dominate for any fixed alpha, so alpha = 0.5 does not by itself implement an equal tradeoff between preferences and AQI. Section 3 reports only qualitative ranking lists and does not provide the raw S values, the ranges of SMF and SAQI, or any normalization step, so the claimed 'balanced' behavior at alpha = 0.5 is not verifiable. The manuscript should either specify the normalization actually used in the app (e.g., min-max, z-score, or rank-based fusion) or report the scale ranges and intermediate scores so that the tradeoff can be assessed.","section":"Section 2 (re-ranking formula)"},{"comment":"The effectiveness claim in Section 4 ('initial experiments highlight the effectiveness') is supported only by qualitative rank lists from two simulated users on synthetic AQI values randomly assigned within [20, 70] over a 1x1 km grid. No quantitative metrics are reported, such as NDCG, precision@k, average AQI of recommended POIs, or preference loss relative to a preference-only ranking. There are no baseline comparisons (e.g., preference-only, AQI-only, or a random re-ranking), no sensitivity analysis over alpha, and no user study. Because the only observed behavior—different alpha values produce different lists—follows directly from the definition of S, the demonstration is closer to a sanity check of the implementation than to evidence of practical effectiveness. Please add at least the measured scores and a baseline comparison, or explicitly reframe the contribution as a system demonstration without an effectiveness claim.","section":"Section 3 (Demonstration Results)"},{"comment":"The paper repeatedly describes AirTOWN as 'privacy-preserving' and lists privacy preservation as one of the system's key pillars, but Section 4 explicitly defers differential privacy to future work and no privacy analysis is provided anywhere in the manuscript. Federated averaging alone does not prevent inference of user information from shared model updates or item-embedding updates, and the statement that 'the Application Layer never shares user data' is not a formal privacy guarantee. This is load-bearing because privacy is one of the four stated design goals and appears in the title and abstract. The authors should either add a concrete threat model and privacy analysis (including the FL update exchange described in Section 2), or soften the privacy claim throughout to something like 'privacy-aware by design' until such an analysis exists.","section":"Title, Abstract, and Section 4 (privacy claim)"}],"minor_comments":[{"comment":"The abstract contains a typo: 'poromsing' should be 'promising'.","section":"Abstract"},{"comment":"The text 'Thispaperintroduces' is missing a space; there are several other spacing issues such as 'a1×1km' in Section 3 that should be corrected in the camera-ready version.","section":"Section 1"},{"comment":"Cork is in Ireland, not the UK; the manuscript says 'Cork, UK' in two places.","section":"Section 1 and Data Resources Layer"},{"comment":"The symbols SMF and SAQI are used in the re-ranking formula but are never explicitly defined beyond the surrounding prose; please define them formally (e.g., as normalized scores from the MF model and the AQI mapping, respectively).","section":"Section 2"},{"comment":"The figure is described as containing red-boxed and blue-boxed lists, but the boxes are not visible in the text version; please annotate the figure clearly so that each panel's alpha value and user identity are legible.","section":"Figure 2"},{"comment":"Reference [13] appears to describe a next-POI recommendation system with geographic and temporal preference, yet it is cited in the introduction as a 'privacy-centric solution'; please verify and correct this citation or reposition it.","section":"References"},{"comment":"The anonymous code link is useful, but the paper would benefit from a short reproducibility statement describing how to run the demonstration, what data are needed, and how the synthetic AQI grid is generated.","section":"Footnote 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads like a demo-track submission: the architecture is sensible and the demonstration scenario is plausible, but the claims outrun the evidence. For a full research track, the authors should either add a small but quantitative evaluation with calibrated scores and baselines, or reframe the paper as an architecture and system description. The title's 'Privacy-Preserving' should also be revisited unless a concrete privacy analysis is added; as it stands, the privacy claim is not supported beyond the use of federated learning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this is a demo paper whose novelty is the specific integration of real-time AQI data with collaborative filtering and federated learning for POI recommendation. I haven't seen that exact combination in prior work, and the architecture is laid out clearly: four layers, MF for personalization, FedAvg with only item embeddings shared, RBF interpolation for sparse sensor coverage. The anonymous code link is a plus. For a demo, the scope is appropriate.\n\nWhat it does well: the authors are honest about limitations. They explicitly defer differential privacy to future work and say a larger user base is needed. The system design is coherent and the motivating problem (health-sensitive users avoiding polluted POIs) is real. The two simulated users, one healthy and one elderly, illustrate how the app could adapt, and the different alpha values do change the lists.\n\nNow the soft spots. The stress-test concern about normalization holds up. The re-ranking score S = alpha*SMF + (1-alpha)*SAQI is the heart of the claimed tradeoff, yet the paper never states how MF preference scores and AQI values are brought to a common scale. MF ratings and AQI readings are not commensurate; unless normalized, alpha=0.5 does not mean equal influence, and the visible ranking changes could be a scale artifact. This is an internal underspecification, not a matter of missing real data. The paper needs to state the normalization step or show the actual S values.\n\nSecond, the evaluation is much weaker than the word \"effectiveness\" suggests. We have two simulated users, synthetic AQI randomly drawn from 20 to 70 on a 1x1 km grid, and no metrics, baselines, or quantitative comparison. The demonstration is a sanity check of the implementation, not evidence that the recommender helps users make healthier choices. For a demo paper that is somewhat forgivable, but the conclusion overreaches.\n\nThe reader's conditional verdict is fair, and I'd go further: the normalization gap is the load-bearing issue. If they fix that and soften the claims, it's a decent demo. If they don't, the central balancing result is not interpretable.\n\nWho is this for: people working on health-aware recommender systems or privacy-preserving POI recommendation. A serious referee from that community could help. Send it to review, expect heavy revision. Not desk-reject material, but close.\n\nRecommendation: referee it, but ask for normalization details, the actual score values, and either real AQI data or at least a sensitivity analysis on the synthetic range.","headline":"New integration, thin evidence: AirTOWN's core re-ranking lacks stated normalization, so the claimed preference/AQI balance is not yet established.","tokens_in":4582,"tokens_out":2146,"would_cite":false,"duration_ms":19278,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A mobile app merges live air-quality readings with personal preferences to recommend healthier nearby places.","keywords":["Air Quality Index","POI recommendation","collaborative filtering","federated learning","privacy","mobile app","re-ranking","health-aware recommender"],"falsifier":"Deploy AirTOWN in a city with real AirSENCE sensor data (e.g., Bari or Cork) for at least a week with a small group of real users, and compare the average AQI of the top-5 recommended POIs under $\\alpha = 0.5$ against the top-5 under $\\alpha = 1$ (preference-only). If the average AQI difference is near zero — or if users rarely visit the AQI-improved suggestions — the central claim of health-conscious re-ranking would fail. A simpler check: measure the spatial AQI variance within a 1 km radius in those cities; if it is tiny, the re-ranking has nothing to act on.","tokens_in":3614,"feed_emoji":"🌍","tokens_out":5753,"duration_ms":47590,"temperature":0.7,"pith_summary":"AirTOWN is a demonstration that a point-of-interest recommender can fold real-time Air Quality Index data into personalized rankings without uploading user ratings to a server. The app uses matrix factorization trained locally on each device, shares only item embeddings through federated learning, and re-ranks candidate places with the weighted score $S = \\alpha S_{MF} + (1-\\alpha) S_{AQI}$. The paper's demonstration simulates two users in Bari with synthetic AQI values on a 1 km grid and shows that varying $\\alpha$ shifts the list from pure AQI ordering to pure preference ordering, with $\\alpha = 0.5$ balancing both. The authors claim this shows the app can adapt to individual health needs, such as an elderly user who prefers lower-pollution spots, while protecting privacy.","feed_headline":"Mobile app re-ranks nearby places by pollution and preference","feed_subtitle":"Federated learning keeps your ratings on-device while AQI data steers recommendations toward cleaner spots.","key_machinery":"The load-bearing identity is the weighted re-ranking score $S = \\alpha \\cdot S_{MF} + (1-\\alpha) \\cdot S_{AQI}$, where $S_{MF}$ is a matrix-factorization preference prediction and $S_{AQI}$ is the Air Quality Index score of the candidate point of interest. The parameter $\\alpha \\in [0,1]$ is the single knob that controls the health-versus-taste trade-off, and the demonstration is primarily about what happens as this knob is turned. The privacy guarantee relies on the federated learning protocol: local user embeddings are trained on-device and never shared, only item embedding updates are aggregated via Federated Averaging, which the paper says keeps raw user data out of the server. For sparse sensor coverage, radial basis function interpolation approximates AQI values at unsensed locations, keeping the re-ranking applicable city-wide.","core_discovery":"On its own terms, the paper claims that AirTOWN integrates real-time pollutant data, personalization, and privacy to offer dynamic, health-conscious recommendations, and that initial experiments highlight the effectiveness of the app in balancing user preferences with air quality. The central mechanism is a client-server recommender where user embeddings are computed locally from survey ratings, item embeddings are aggregated on the server through Federated Averaging, and a final re-ranking $S = \\alpha \\cdot S_{MF} + (1-\\alpha) \\cdot S_{AQI}$ lets a single parameter interpolate between preference-only and AQI-only ordering. The demonstration with two simulated users in Aldo Moro Square, Bari — one healthy, one elderly with increased sensitivity — shows qualitatively distinct lists as $\\alpha$ is set to $0$, $1$, $0.5$, and $0.3$, which the authors interpret as evidence of adaptability. Because no real user data leaves the device, the paper argues this addresses the privacy gap in existing health-aware recommenders.","pith_inferences":["The paper's effectiveness claim is supported only by a two-user simulation with randomly assigned AQI values; a real deployment would need to show that actual AirSENCE data exhibits enough AQI variation within a 1 km radius for re-ranking to change user choices.","Sharing item embeddings, as the design does, still carries privacy risk in federated learning, so the stated future work on differential privacy is essential if the privacy guarantee is to be stronger than the current architecture.","The re-ranking formula implicitly assumes $S_{MF}$ and $S_{AQI}$ are on comparable numeric scales; the paper does not describe normalization, so in practice the effective weight of AQI could differ from $\\alpha$.","A testable prediction of the design is that users will deviate toward lower-AQI POIs when $\\alpha$ is lowered; this could be verified in a controlled field study with real pollution data and measured route choices."],"forward_implications":["Users in cities with sensor coverage could receive POI rankings that respond to pollution spikes in near real time, while their individual ratings and location history stay on-device.","The single parameter $\\alpha$ gives a natural control for health sensitivity: an app could set it lower for elderly or asthmatic users and higher for users who prioritize taste.","The architecture extends to any environmental factor beyond AQI — noise, heat, pollen — by swapping the second score in the re-ranking.","If real-sensor demonstrations confirm the behavior seen with synthetic data, the approach could be a lightweight alternative to deep learning recommenders in privacy-sensitive domains."],"supporting_citations":[{"why":"Supplies the WHO air-pollution statistic that motivates the health problem AirTOWN addresses.","marker":"[1]"},{"why":"Baseline that monitors ambient air pollution exposure but does not suggest healthier POIs.","marker":"[14]"},{"why":"Baseline that monitors and predicts pollution along routes without offering POI recommendations.","marker":"[10]"},{"why":"Health-aware food recommender relying on static health metrics, contrasting with AirTOWN's real-time environmental input.","marker":"[12]"},{"why":"Example of health nudging in recommender systems that does not incorporate live air quality.","marker":"[11]"},{"why":"Real-time POI recommender that integrates geography and temporal preference but lacks environmental awareness.","marker":"[13]"},{"why":"Privacy-preserving POI recommendation work that AirTOWN contrasts with because it lacks real-time environmental integration.","marker":"[3]"}],"fun_headline_variants":["Pollution-aware app suggests cleaner spots, keeps data private","App blends air quality and taste to rank nearby places","Federated learning powers pollution-friendly POI recommender","AirTOWN: privacy-first app steers users to cleaner venues","Tunable ranking mixes user preference with real-time AQI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The demonstration's central assumption is that AQI values randomly assigned between 20 and 70 on a simulated 1 km grid, plus two scripted users, represent real urban air quality and real user behaviour; if actual sensor data has different spatial patterns, noise, or gaps, the observed re-ranking behaviour may not occur in a real city.","fun_headline_variants_meta":{"raw":{"variants":["Pollution-aware app suggests cleaner spots, keeps data private","App blends air quality and taste to rank nearby places","Federated learning powers pollution-friendly POI recommender","AirTOWN: privacy-first app steers users to cleaner venues","Tunable ranking mixes user preference with real-time AQI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000739,"raw_usage":{"total_tokens":3270,"prompt_tokens":882,"completion_tokens":2388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":2305}},"tokens_in":498,"tokens_out":2388,"duration_ms":14887,"temperature":1.0,"reasoning_tokens":2305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:45:53.546937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy AirTOWN in a city with real AirSENCE sensor data (e.g., Bari or Cork) for at least a week with a small group of real users, and compare the average AQI of the top-5 recommended POIs under $\\alpha = 0.5$ against the top-5 under $\\alpha = 1$ (preference-only). If the average AQI difference is near zero — or if users rarely visit the AQI-improved suggestions — the central claim of health-conscious re-ranking would fail. A simpler check: measure the spatial AQI variance within a 1 km radius in those cities; if it is tiny, the re-ranking has nothing to act on.","supporting_citations":[{"cited_title":"https://www.who.int/health-topics/air-pollution (Accessed June 16, 2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the WHO air-pollution statistic that motivates the health problem AirTOWN addresses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Privacy-preserving POI recommendation work that AirTOWN contrasts with because it lacks real-time environmental integration."}],"review_version":1}