{"id":"91579678-0a25-4a94-ae70-e256bdd1b887","arxiv_id":"1908.05544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A proximity-based gossip recommender design is introduced; the Android prototype can exchange ratings within about 6 meters, but the filtering and recommendation steps are not implemented.","lead":"This short paper proposes Propagate and Filter, a method for smartphones to swap movie ratings directly with nearby devices over short-range wireless links. The authors measure an Android prototype and report about 26 seconds average pairing delay, reliable transmission within 6 meters, and roughly 5-6% hourly battery drain while advertising.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline feasibility claim is not directly measured: the 25.9 s delay comes from a 1 m 'guaranteed connectivity' trial, while the 70–80% success rates at 6 m carry no timing data, so 'reliable within 6 m' conflates two separate experiments.","rationale":"The reader and I agree that the paper should remain CONDITIONAL, but the weakest point differs. The reader emphasizes the placeholder sampling and recommendation algorithms as the main load-bearing gap; I think the more direct gap is that the headline feasibility figure is not an integrated measurement. The paper's central claim is explicitly technical feasibility, not recommendation quality, so a flaw in the propagation metrics damages the central claim more directly than the placeholder algorithms do. The suggested re-measurement is cheap and would settle whether the '25.9 s / 6 m' claim holds jointly. If it does not, the conclusion should qualify the stated use cases; if it does, the CONDITIONAL verdict is unchanged. Because the reader already marked the paper CONDITIONAL and my concern does not refute the existence of a working short-range prototype, I keep the verdict at UNCHANGED rather than moving to REJECT. The 'partial' agreement reflects that the reader's rationale did note missing sample counts and error bars, though the stated weakest assumption concerned privacy-preserving similarity and recommendation convergence rather than the feasibility metric itself.","tokens_in":8749,"tokens_out":6160,"duration_ms":62814,"concrete_test":"Rerun the propagation experiment with N ≥ 30 per condition at distances 3 m, 6 m, and 10 m, with and without obstacles, recording time from the start of discovery until the full rating list is exchanged or the attempt times out; report median, interquartile range, and success fraction per condition. Then compute the probability of completing an exchange within T = 30, 60, and 120 s and compare these against encounter-duration distributions for traffic lights and public transport. If the 6 m median delay is close to 25.9 s and more than 80% of 6 m attempts finish within 60 s, the feasibility claim is supported; otherwise the abstract's joint claim should be revised or restricted to the 1–3 m regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is technical feasibility of the propagation step (abstract; Section 5 conclusion), but the supporting evidence is split across two experiments with mismatched conditions. Experiment (5) reports an average initial connection delay of 25.9 s (min 11 s, max 41 s), yet the devices were placed at 1 m 'in order to guarantee connectivity'—not at the 3–6 m distances where the success-rate claims are made. Experiment (4) reports success rates at 3, 6, 10, and 12 m (Table 1), but records only whether a connection eventually succeeded; it does not report how long those connections took, how many attempts timed out, or the timeout threshold. Thus the abstract's phrase 'average ad-hoc connection delays of 25.9 seconds and reliable connection success rates within 6 meters' assembles a complete 6 m feasibility statement from two independent measurements, and the joint property—a 6 m encounter completes an exchange within an acceptable time—is never measured. At 6 m with obstacles the success rate is already only 70%; if the delay at that distance is also longer than 25.9 s, plausibly due to weaker signal and retries, the stated use cases (waiting at traffic lights or public transport) may fail for a large fraction of encounters. No confidence intervals or per-condition sample sizes are given, and the min/max range 11–41 s is too wide to infer a reliable upper bound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Propagate and Filter, a gossip-based method for exchanging item preference data between smartphones in physical proximity, intended to address data scarcity and privacy in pervasive recommender systems. It describes four data types (peer preference list, neighborhood preference list, similarity data, context data), a propagation step, and a filtering step, then evaluates an Android prototype that implements only the propagation step using Google's Nearby Connections API. The experiments report lossless transfer of 1000 ratings, multi-device transfer, transfer in public transportation, connection success rates by distance (Table 1), an average initial connection delay of 25.9 s measured at 1 m, and pre-connection battery drain of about 5.8%/h. The authors conclude that the propagation step is technically feasible for scenarios such as waiting at traffic lights or taking public transportation, while acknowledging limitations in range and battery consumption.","tokens_in":9024,"tokens_out":4802,"duration_ms":47447,"significance":"The paper's measured results provide useful early empirical evidence that opportunistic proximity-based propagation of rating data is feasible without internet connectivity, and the authors are commendably explicit about limitations such as the 6 m effective range and the battery cost of continuous advertising and discovery. The work is explicitly preliminary: the Filter step and the recommendation algorithm are placeholders, so the claimed data-scarcity and privacy benefits are not empirically demonstrated. The paper does not offer machine-checked proofs or parameter-free derivations; its value is as a system design plus an empirical feasibility study of the propagation step. If the propagation-step scope is made precise, the results are a reasonable contribution for a systems-oriented venue.","major_comments":[{"comment":"The abstract's feasibility statement combines a delay measurement taken at 1 m \"in order to guarantee connectivity\" (Experiment 5) with success rates measured at 3, 6, 10, and 12 m (Table 1) for which no timing information is reported. The joint claim that a 6 m encounter completes an exchange within an acceptable time is therefore not directly supported. At 6 m with obstacles the success rate is already 70%; if connection setup at that distance takes longer than the 25.9 s mean measured at 1 m, the stated use cases (traffic lights, public transport) may be optimistic. Please report delay statistics and timeout thresholds for the distances in Table 1, or explicitly scope the delay claim to the 1 m condition.","section":"Section 5, Experiments (4) and (5); Section 5.1"},{"comment":"No per-condition sample sizes, confidence intervals, or timeout thresholds are reported for any of the six experiments. The text says experiments were re-run 10 times by default, but it is not stated which conditions used different numbers of runs, nor whether the differences between, say, 80% and 70% success at 6 m are meaningful given the sample size. Because the central feasibility conclusion rests on these rates and on the 25.9 s mean delay, the paper should report N, variation (min/max or confidence intervals), and a precise definition of a failed attempt.","section":"Section 5, Table 1 and Experiments (1)-(6)"},{"comment":"The Filter step (similarity comparison, neighborhood preference list resampling) and the recommendation algorithm are explicitly left as placeholders, so the claim in Section 3.5 that Propagate and Filter \"addresses the profile data scarcity problem\" is not supported by the evaluation. The abstract's phrase \"the approach's technical feasibility\" should be scoped to the propagation step; as written, it invites the reader to infer end-to-end recommender feasibility. If the intended contribution is limited to propagation, a one-sentence scope clarification would suffice; otherwise, evidence about the sampling and recommendation steps is needed.","section":"Sections 3.3, 3.5, and 6"}],"minor_comments":[{"comment":"The bullet list says the paper presents \"an implementation of the propagation part\" and \"its evaluation,\" which is accurate, but the title and abstract emphasize the full Propagate and Filter method; consider aligning the framing so the reader immediately sees that the evaluation is only of propagation.","section":"Section 1, final paragraph"},{"comment":"The phrase \"above the k-th highest\" is ambiguous; it should say something like \"above the k-th highest similarity among known peers\" or \"in the top k by similarity.\"","section":"Section 3.3, step (2)"},{"comment":"The captions should state the number of runs per condition and the units (e.g., percentage, %/h) directly, rather than requiring the reader to infer them from the text.","section":"Table 1 and Table 2 captions"},{"comment":"The phrase \"as it is commonly the case in underground trains\" reads awkwardly; consider rewording to \"as is common in underground trains.\"","section":"Section 4.2"},{"comment":"The phrase \"raise, communicate, and process data\" appears to contain a typo; \"raise\" should likely be \"collect\" or \"capture.\"","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"This is a short, systems-oriented conference paper being evaluated in a journal context. The main barrier to acceptance is the mismatch between the 1 m delay measurement and the 6 m success-rate claim, plus the absence of statistical detail; both are fixable. The placeholder Filter step is acceptable only if the authors explicitly scope the feasibility claim to propagation. I would not reject on the basis of the untested recommendation algorithms, provided the revision makes the scope unambiguous."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"One thing to know: the paper's headline feasibility claim is not directly measured. The 25.9-second average connection delay comes from an experiment at 1 m to 'guarantee connectivity,' while the 70–80% success rates at 6 m come from a separate experiment that only records whether a connection succeeded, not how long it took. The abstract stitches these together into 'reliable connection success rates within 6 meters,' but the joint property—a 6 m encounter completing an exchange within an acceptable time—is never measured. At 6 m with obstacles the success rate is already 70%; if the delay there is also longer than 25.9 s, the traffic-light and public-transit scenarios get shaky.\n\nThe paper does what a feasibility study should otherwise do. It proposes a named method, Propagate and Filter, that transplants gossip-based decentralized recommender ideas to proximity settings. It builds a real Android prototype on the Nearby Connections API and measures genuine things: 1,000 ratings transferred losslessly, multi-device exchange, public-transit operation, and battery drain around 5.8%/h pre-connection. The authors are explicit that sampling and recommendation algorithms are placeholders and label the results preliminary. That honesty is real.\n\nThe soft spots beyond the measurement mismatch: no per-condition sample sizes or confidence intervals, no timeout threshold for a failed attempt, and the app ships as a binary without source or raw data, so the numbers can't be reproduced. The abstract's privacy and data-scarcity claims outrun what the paper shows. Privacy by disconnection only hides connection metadata; the propagated neighborhood preference lists could still leak content. And the resampling idea—mixing lists from the k most similar peers to approximate latent communities—is plausible but untested. These are the usual gaps of a feasibility paper, not fatal flaws.\n\nRead this if you work on proximity-based or decentralized recommender systems. It's a short paper, useful as engineering evidence about what the Nearby Connections API can do. The citation pattern is fine. I'd send it to referees: a competent reviewer could catch the conflation and ask for a joint delay-at-distance experiment, which is exactly what the paper needs. I'd cite it only for the transport-layer numbers, not for any recommendation-quality claim.","headline":"A feasible propagation step, but the paper's 'reliable within 6 meters' headline conflates a 1 m delay measurement with separate 6 m success rates—worth a referee to fix.","tokens_in":9559,"tokens_out":5648,"would_cite":true,"duration_ms":47023,"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":"Nearby smartphones can exchange ratings reliably and offline, a prototype claims.","keywords":["opportunistic network","decentralized recommender system","pervasive recommender system","gossip protocol","data scarcity","data privacy","information dissemination","proximity-based mobile social network"],"falsifier":"Run the currently placeholder sampling and recommendation steps on real rating histories: if pairs selected as similar by privacy-preserving similarity data share no more preferred items than random pairs, or if resampling from their mixed lists does not improve recommendation accuracy over no exchange, the approach's benefit collapses. For the propagation half, measure connection delay and success rate in dense pedestrian traffic: if the average setup time exceeds the typical duration of street encounters, the stated use cases shrink to waiting and transit only.","tokens_in":8550,"feed_emoji":"📱","tokens_out":6077,"duration_ms":56831,"temperature":0.7,"pith_summary":"This paper argues that recommender systems can work without a central server or internet connection by having nearby smartphones gossip item ratings directly to one another. It proposes Propagate and Filter, a two-step method in which devices exchange privacy-preserving similarity data and aggregated preference lists whenever they are in physical proximity, then keep only data from the most similar peers. The paper reports a prototype Android app that exchanged 1,000 ratings losslessly, connected with 100% success at 3 meters and 70-80% at 6 meters, and took 25.9 seconds on average to establish a connection. On that evidence it claims the propagation step is technically feasible for settings such as waiting at traffic lights or riding public transport.","feed_headline":"Nearby phones swapped 1,000 ratings without data loss in prototype","feed_subtitle":"A 25.9-second average connection delay and 70-100% success within 6 meters back offline recommender gossip.","key_machinery":"The central object is 'Propagate and Filter', a two-phase method. In the propagate phase, nearby devices exchange two kinds of data: similarity data, chosen to be privacy-preserving, and a neighborhood preference list, an aggregated list of item ratings from the k most similar previously encountered peers. In the filter phase, a device compares similarity data with the sender, and if the sender ranks among the k most similar peers, resamples its own neighborhood preference list from the sender's list and those of the other similar peers; recommendations are then computed locally from all four data pools (peer list, neighborhood list, similarity data, context data). The mechanism is designed so that recommendation flows between similar peers and dries out between dissimilar ones, emulating gossip-protocol convergence without needing a persistent network overlay.","core_discovery":"The paper's central claim is that the propagation step of Propagate and Filter is technically feasible today: a smartphone can act as a node in an ephemeral, offline recommender network, passing along ratings and similarity data during brief encounters. The prototype transmitted all 1,000 ratings without loss, handled simultaneous exchanges among four devices, worked in Berlin buses and underground trains without internet, and achieved reliable connections within 3-6 meters at an average initial connection delay of 25.9 seconds. The authors present this as preliminary evidence that pervasive recommender systems can address both data scarcity and some privacy risks by exchanging data only in immediate proximity and then disconnecting.","pith_inferences":["The recommendation-quality layer is untested: if privacy-preserving similarity data does not actually select peers with correlated tastes, mixing their lists could dilute rather than sharpen local preferences; a field trial with ratings and follow-up choices would settle this.","The 25.9-second connection delay is an average over one API and phone generation; newer radios and connection caching could plausibly push delay below the duration of a pedestrian pass, opening street-level encounters the paper currently excludes.","The same propagate-and-filter loop could double as a context-discovery mechanism: because encounters are tagged with location and time, exchanged lists implicitly encode place-based tastes, which might enable location-aware recommendations without a central profile.","Because the prototype exchanged only peer preference lists rather than the full four-layer data stack, the measured feasibility does not yet cover similarity-data propagation or the cost of the filtering computation; those remain open."],"forward_implications":["If propagation is feasible, recommendation models can be built on-device with no internet: ratings arrive during physical encounters and movie metadata is fetched later when connectivity returns.","Data scarcity in pervasive recommenders can be reduced because recommendations relay between peers that are never geographically close, through chains of similar peers.","Privacy by disconnection follows: no persistent network overlay exists, queries cannot be issued at will, and a peer is invisible except during an active encounter; only the data a peer chooses to share is exposed, and only at contact time.","The measured constraints bound the use cases: 25.9 seconds average setup and 6-meter range fit waiting-at-traffic-lights and public-transport settings, while 70-80% success at 6 meters and roughly 5% per hour battery drain set practical limits."],"supporting_citations":[{"why":"Closest prior work proposing device-to-device raw profile exchange in an opportunistic network, which Propagate and Filter extends by addressing privacy and scalability.","marker":"[25]"},{"why":"Gossip-based peer sampling supplies the conceptual mechanism for converging similar peers that Propagate and Filter adapts to spatial movement.","marker":"[19]"},{"why":"T-Man gossip overlay topology management is cited as the protocol that normally retains connectivity and that spatial peers cannot use directly.","marker":"[18]"},{"why":"The authors' own earlier mobile software architecture is the base on which the prototype's propagation step is built.","marker":"[5]"},{"why":"Similarity estimation on item vectors in proximity-based mobile social networks is cited as a privacy-preserving route for the comparison step.","marker":"[4]"},{"why":"Latent user similarity on texting data is cited as an alternative privacy-preserving similarity signal.","marker":"[15]"},{"why":"Proximity at a music festival as a reliable similarity estimate motivates using spatial encounters as peer sampling.","marker":"[11]"},{"why":"Information dissemination in urban areas of interest via opportunistic networks backs the claim that urban mobility supplies enough encounters.","marker":"[9]"},{"why":"Gossip overlays for self-emerging user communities show the decentralized-recommender tradition that this paper moves into pervasive settings.","marker":"[3]"}],"fun_headline_variants":["Phones gossip ratings offline, 25.9s hookup","No internet? Phones still swap 1000 ratings","Proximity recommender: gossip with 100% success","Offline recommender gossip passes 1000-item test","Phones whisper ratings in 25.9s, no cloud needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"If similarity data fails to pick out genuinely similar peers, or if mixed neighborhood lists carry little signal, then exchanging ratings with strangers yields poor recommendations and the claimed benefits over centralized recommenders disappear.","fun_headline_variants_meta":{"raw":{"variants":["Phones gossip ratings offline, 25.9s hookup","No internet? Phones still swap 1000 ratings","Proximity recommender: gossip with 100% success","Offline recommender gossip passes 1000-item test","Phones whisper ratings in 25.9s, no cloud needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1177,"prompt_tokens":893,"completion_tokens":284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":509,"tokens_out":284,"duration_ms":3189,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:09:57.817473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the currently placeholder sampling and recommendation steps on real rating histories: if pairs selected as similar by privacy-preserving similarity data share no more preferred items than random pairs, or if resampling from their mixed lists does not improve recommendation accuracy over no exchange, the approach's benefit collapses. For the propagation half, measure connection delay and success rate in dense pedestrian traffic: if the average setup time exceeds the typical duration of street encounters, the stated use cases shrink to waiting and transit only.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gossip-based peer sampling supplies the conceptual mechanism for converging similar peers that Propagate and Filter adapts to spatial movement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"T-Man gossip overlay topology management is cited as the protocol that normally retains connectivity and that spatial peers cannot use directly."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The authors' own earlier mobile software architecture is the base on which the prototype's propagation step is built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Similarity estimation on item vectors in proximity-based mobile social networks is cited as a privacy-preserving route for the comparison step."},{"cited_title":"affinity: A System for Latent User Similarity Comparison on Texting Data","cited_arxiv_id":"1904.01897","evidence_quote":"Latent user similarity on texting data is cited as an alternative privacy-preserving similarity signal."},{"cited_title":"Norrie, and Michael Grossniklaus","cited_arxiv_id":null,"evidence_quote":"Proximity at a music festival as a reliable similarity estimate motivates using spatial encounters as peer sampling."},{"cited_title":"Chancay-GarcÃŋa, E","cited_arxiv_id":null,"evidence_quote":"Information dissemination in urban areas of interest via opportunistic networks backs the claim that urban mobility supplies enough encounters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gossip overlays for self-emerging user communities show the decentralized-recommender tradition that this paper moves into pervasive settings."}],"review_version":1}