{"id":"2daa9656-4477-4317-be9e-614e92514aac","arxiv_id":"1908.08929","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using cosine similarity of Wi-Fi scans, the authors cluster visits into indoor points of interest and detect common places among users in a shopping mall.","lead":"This paper tests whether phone Wi-Fi scans can reveal which indoor places people visit. Crowdsensing data from a shopping mall shows that clustering similar Wi-Fi patterns can pick out shared points of interest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is supported only by anecdotal tables; no quantitative accuracy or held-out evaluation is reported, and the 0.5 cosine threshold is tuned to the displayed data, so the claimed POI-identification capability is not established beyond a demonstration.","rationale":"The central assertion is a feasibility claim: the system can identify a user's POIs and shared POIs across users. For that to be true, the similarity function in Equations 6-9 must make intra-POI scan pairs more similar than inter-POI scan pairs under a single threshold, despite missing APs and heterogeneous devices. The paper's experimental section contains exactly the material needed to test this, but no test is reported. Table II compares two thresholds on a single day and states that 0.5 matches ground truth without giving counts or scores; Table III shows colored repeats for one user; the mall experiment in Section IV-C maps detected communities but provides no accuracy measure. The threshold itself is justified by the same table used as evidence, so those displays are not an independent validation. This lack of quantitative validation is more fundamental than the device-calibration issue flagged by the reader: even if all devices were identical, the paper would still not show how often the 0.5 threshold returns correct POI labels. The proposed leave-one-user-out test uses data that already exists, so it is feasible and would directly measure the separation. If it passes, the central claim gains real support; if it fails or shows strong device dependence, the paper needs recalibration or a device-aware similarity model. Until such a check is reported, the claim should remain conditional rather than accepted as demonstrated.","tokens_in":6771,"tokens_out":5349,"duration_ms":59392,"concrete_test":"Using the existing 11-user mall dataset (Table IV) with ground-truth POI labels, run a leave-one-user-out evaluation: for each user, build POI fingerprints from the other users' data, then classify every ground-truth labeled stay by the held-out user as a revisit to a known POI if the cosine similarity to the best-matching fingerprint is at least 0.5. Report precision, recall, and F1 per user and per phone model. If same-POI cross-device pairs do not clearly separate from different-POI pairs, or if F1 varies strongly by phone model, the fixed 0.5 threshold and uncalibrated RSS comparison do not support the paper's central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that Wi-Fi cosine similarity identifies distinct POIs and common POIs across users requires that, at a fixed threshold (epsilon = 0.5, Section III-A), same-POI scan pairs score systematically higher than different-POI scan pairs, including pairs from different phone models (Table IV). The manuscript never measures this separation. Table II shows one user's day at two threshold values and asserts, without a count or a score, that 0.5 matches the ground truth; Table III shows repeated POI IDs for one user; Section IV-C reports Louvain modularity for thresholds but no precision or recall against the mall ground truth. In addition, epsilon is selected using the same displayed examples that are then presented as supporting evidence, so those examples are not an independent test. Consequently the central feasibility claim is under-supported: one anecdote shows plausible behavior, but there is no quantitative basis for expecting the pipeline to work on new users, new devices, or new buildings.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a mobile crowdsensing system for indoor point-of-interest (POI) identification. A smartphone application collects Wi-Fi scans (MAC addresses and RSS values) every five minutes, compresses them with gzip, and uploads them to a backend. The backend clusters scans using a modified DBSCAN with cosine similarity of RSS vectors, assigns a POI fingerprint to each cluster, and then uses Louvain community detection on pairwise POI similarities to find common POIs across users. The experimental section reports results for two similarity thresholds for one user's day, for one user's revisited POIs, and for common POIs among 11 users in a shopping mall. The paper's central claim is that this pipeline can identify distinct and common indoor POIs.","tokens_in":6975,"tokens_out":5977,"duration_ms":55611,"significance":"If established, the approach would offer a low-power, infrastructure-light way to infer indoor places from ordinary smartphone Wi-Fi scans, with applications in mobility and behavior analysis. The strengths of the paper are its real deployment with volunteer users and ground-truth labels, the clear system architecture, and the explicit treatment of data compression. However, the evidence presented is anecdotal and lacks quantitative accuracy assessment; the central similarity threshold is selected on the same data used for validation. Consequently, the current manuscript does not substantiate the abstract's claim that the system 'can identify' POIs beyond a demonstration.","major_comments":[{"comment":"The threshold ε=0.5 is introduced in Section III-A with the justification 'based on experimental evaluation as shown in the Table II', and Table II.A is then presented as the matching case. Because ε is chosen by checking which value reproduces the ground-truth labels in that same table, the table is not an independent test. There is no held-out evaluation on other days, users, or environments; every later result, including the community detection in Section IV-C, uses this same fitted threshold. This circularity makes the central feasibility claim unsupported.","section":"III-A, Table II"},{"comment":"No quantitative accuracy metric (precision, recall, F1, or a confusion count) is reported for POI extraction. Tables II and III are anecdotal label sequences, and the shopping-mall common-POI experiment in Section IV-C lacks ground-truth validation: Table V reports only Louvain modularity, and the authors select ε=0.5 despite the higher modularity at ε=0.3 (0.803 vs 0.692), giving no label-based justification for that choice. The claim that the system identifies common POIs across users is therefore not measured.","section":"IV-B, IV-C"},{"comment":"The cosine similarity in Equations (6)-(9) is computed only over MAC addresses common to the two fingerprints. No mechanism is described for missing APs, RSS normalization, or device-specific calibration, even though Table IV lists 11 different smartphone models. Intermittent Wi-Fi scans can thus produce same-POI pairs with few common MACs and deflated scores, while pairs that happen to share a small set of MACs can receive high scores. The cross-user common-POI result relies directly on this unvalidated comparability assumption.","section":"III-B, Eqs. (6)-(9)"},{"comment":"The pseudocode of Algorithm 1 is not a well-defined procedure as printed: the loop variable z1 is incremented only in the else branch, so after a cluster is formed the same α may be revisited indefinitely, and the inner loop's control flow for β ∈ Vp is ambiguous. This makes the core clustering algorithm not fully reproducible from the paper.","section":"Algorithm 1"}],"minor_comments":[{"comment":"The last paragraph says 'Section IV concludes the paper with a discussion and future work'; the conclusion is actually Section V.","section":"Section I"},{"comment":"The text refers to POI IDs highlighted in the same color, but the table as presented does not use color; please use distinct markers or labels to indicate revisits.","section":"Table III"},{"comment":"The line 'N = 0' should initialize N to the empty set rather than the integer 0.","section":"Algorithm 2"},{"comment":"The paper does not report the number of Wi-Fi scans per user or the total duration of the shopping-mall data collection, which would help assess the coverage and stability of the reported results.","section":"IV"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is currently below the bar for this journal because the central claim is supported by anecdotal examples and a circularly chosen threshold. The real data collection and the proposed pipeline are potentially useful, and the weaknesses are addressable by adding a validation protocol: split data into tuning/held-out sets, report precision/recall/F1 for individual and common POI extraction, and validate the mall result against independent labels or a reproducible code release. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:1908.08929. It's a system paper: an Android app scans Wi-Fi every five minutes, DBSCAN with cosine similarity clusters the scans into POIs, and Louvain community detection finds POIs shared across users. The experimental component is real—eleven volunteers, multiple phone models (Table IV), ground-truth labels for a home/office day, and a shopping-mall deployment. The compression numbers in Table I are a useful engineering detail. This is genuinely a working pipeline, not just a simulation.\n\nWhat's new is modest. The underlying observation—Wi-Fi RSS fingerprints distinguish indoor locations—dates back to the indoor-localization literature (their refs [13]–[15]). The specific contribution is applying that to low-sampling crowdsensing without manual fingerprinting, which extends their own earlier PerCom workshop paper. That's a legitimate incremental step, but not a conceptual leap.\n\nThe soft spots are real and, in my reading, exactly where the stress-test note points. The central evaluation is anecdotal. Table II shows one user's day at two thresholds and asserts that 0.5 matches ground truth; no counts, no precision/recall, no held-out data. The 0.5 threshold is selected because of that same table, so the examples are not an independent test. The shopping-mall common-POI result has no ground-truth validation—Figure 3 shows maps, but there's no check against where users actually went. And the similarity metric in Equations (6)-(9) sums only over common MAC addresses, which means device model differences are unaddressed. With eleven phone models in Table IV, that's a significant concern.\n\nI'd call the soundness weak-to-moderate, not broken. The pipeline is plausible and the presentation is honest—they show the 0.6 threshold over-segmenting and report modularity scores even when they don't select the max. What's missing is a proper quantitative evaluation.\n\nFor a peer-review decision: I'd send it out. A serious referee can ask for the missing metrics and a proper train/test split, and the experimental effort deserves that engagement. I would not cite it as it stands. Reading group? Maybe, as a case study in how to (and how not to) evaluate a crowdsensing system.","headline":"A real crowdsensing POI pipeline undone by an anecdotal, threshold-tuned evaluation; worth a referee but not a citation.","tokens_in":7472,"tokens_out":3866,"would_cite":false,"duration_ms":38219,"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":"Indoor places a person visits, and places shared across people, can be identified from ordinary smartphone Wi-Fi scans by clustering RSS vectors with cosine similarity.","keywords":["indoor points of interest","mobile crowdsensing","Wi-Fi fingerprinting","cosine similarity","DBSCAN clustering","community detection","Louvain method","received signal strength"],"falsifier":"Take two of the phone models in Table IV (for example, OnePlus 3 and Sony Xperia Z3), leave them side by side in one room for the 20 minutes the paper uses to define a POI, and compute the pairwise cosine similarities of their five-minute scans: if a large share of same-location cross-device pairs falls below the 0.5 threshold, or same-location similarity is not separated from different-location similarity, the claimed identification collapses.","tokens_in":6612,"feed_emoji":"📶","tokens_out":9574,"duration_ms":96286,"temperature":0.7,"pith_summary":"This paper tries to establish that ordinary Wi-Fi scans from a smartphone are enough to identify the distinct indoor places a person visits and the places different people visit in common. The authors collect scans every five minutes, represent each scan as a set of access-point MAC addresses with received signal strengths, and use cosine similarity between those RSS vectors as the distance in a modified DBSCAN clustering algorithm. With a similarity threshold of 0.5, the clusters match ground-truth labels such as home, office, meeting room, and canteen, and repeated visits to the same place fall into the same cluster. They then feed pairwise similarities between per-user POI fingerprints into Louvain community detection and report that common POIs across 11 users in a shopping mall emerge correctly. If the claim holds, indoor place recognition becomes a by-product of existing phone hardware, since the only input is the phone's routine Wi-Fi scan and no GPS accuracy is required.","feed_headline":"Wi-Fi scans reveal indoor spots people visit and share","feed_subtitle":"Five-minute Wi-Fi scans, cosine-scored and clustered, matched real indoor places for one user and across 11 shoppers.","key_machinery":"The load-bearing identity is the cosine similarity of two RSS vectors, $C = Y/(\\sqrt{d_1}\\sqrt{d_2})$, where $Y$ is the dot product taken over access-point MAC addresses appearing in both scans and $d_1,d_2$ are the self-dot-products of each scan. Since measured RSS values are negative dBm numbers, scans from one place point in similar directions in RSS space, making $C$ close to 1 even when absolute signal strengths differ; the paper uses $C \\geq 0.5$ as the DBSCAN neighbourhood test and sets minPts $= 4$ because a POI is defined as a stay of at least 20 minutes at a 5-minute scan cadence. The same pairwise $C$ values become edge weights in a Louvain community-detection graph, whose modularity optimization separates shared POIs from private ones.","core_discovery":"The central claim is that a Wi-Fi RSS fingerprint encodes a place well enough to support indoor point-of-interest extraction from low-rate crowdsensed data. Concretely, the paper converts each five-minute scan into a vector of RSS values keyed by access-point MAC addresses, clusters the day's scans with DBSCAN in which the neighbourhood relation is 'cosine similarity at least 0.5' and a cluster needs four scans, and reports that the resulting clusters match the ground-truth POIs of a user's day, including revisits to the same POI. The threshold comparison in Table II is the key evidence: at 0.5 the extracted sequence is exactly Home, Office, Meeting room, Canteen, Office, Home, while at 0.6 an extra POI appears because small environmental changes alter the RSS. For the common-POI scenario, 41 detected POIs across 11 shoppers are connected by 820 pairwise similarities and partitioned by Louvain community detection, yielding communities that correspond to places such as the food court and shops in a mall.","pith_inferences":["The paper does not compare same-location similarity across the 11 phone models directly; a reader should expect that cross-device RSS offsets and different antenna gains could lower cosine scores, so per-device calibration or a lower threshold for cross-user matching is a testable refinement.","The five-minute scan cadence and 20-minute minimum stay set a floor on what counts as a POI; shorter encounters such as a coffee counter or ATM will be invisible, so applications aimed at those places would need a faster scan rate and would pay a battery cost.","The same pairwise-similarity graph used for community detection can be reused for recommendation: POIs that recur inside one community are the places a new user with overlapping patterns is most likely to visit next, which the paper mentions only as future profiling work."],"forward_implications":["A phone carried through a day can produce a start/end time-stamped list of the indoor places its owner visits, using only its own five-minute Wi-Fi scans.","When the same user returns to a previously visited POI, the cosine similarity of the new scans to the stored POI fingerprint puts the revisit in the same cluster, enabling visit counts and dwell times.","Across a group of users, the same similarity threshold of 0.5, applied to pairwise POI similarities and fed into Louvain community detection, yields the shared POIs of the environment.","The six-hour compressed upload size is a few kilobytes, so the sensing burden on the network is small enough for the proposed crowdsensing workflow."],"supporting_citations":[{"why":"Selects DBSCAN over other clustering algorithms for low-sampling crowd-sensed smartphone sensor data, motivating the paper's modified DBSCAN approach.","marker":"[10]"},{"why":"Sets the five-minute Wi-Fi scan interval as the battery-conscious standard in Android, which fixes the scan cadence and the minPts=4 parameter.","marker":"[18]"},{"why":"Defines the original DBSCAN density-based clustering algorithm whose distance metric the paper replaces with cosine similarity.","marker":"[20]"},{"why":"Supplies the Louvain modularity-optimization algorithm used to detect communities of POIs shared across users.","marker":"[21]"}],"fun_headline_variants":["Wi-Fi crowdsensing maps indoor visits in minutes","Cosine-scored Wi-Fi scans reveal indoor hangouts","Short Wi-Fi probes identify indoor points of interest","Indoor POIs from Wi-Fi fingerprints: an experiment","Crowdsensed Wi-Fi finds shared indoor locations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that RSS measurements from different smartphone models can be compared directly with cosine similarity, and that access points missing from one scan can simply be left out of the dot product, without any device calibration or minimum-overlap rule.","fun_headline_variants_meta":{"raw":{"variants":["Wi-Fi crowdsensing maps indoor visits in minutes","Cosine-scored Wi-Fi scans reveal indoor hangouts","Short Wi-Fi probes identify indoor points of interest","Indoor POIs from Wi-Fi fingerprints: an experiment","Crowdsensed Wi-Fi finds shared indoor locations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2776,"prompt_tokens":857,"completion_tokens":1919,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":1844}},"tokens_in":473,"tokens_out":1919,"duration_ms":14301,"temperature":1.0,"reasoning_tokens":1844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:22:22.867743+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two of the phone models in Table IV (for example, OnePlus 3 and Sony Xperia Z3), leave them side by side in one room for the 20 minutes the paper uses to define a POI, and compute the pairwise cosine similarities of their five-minute scans: if a large share of same-location cross-device pairs falls below the 0.5 threshold, or same-location similarity is not separated from different-location similarity, the claimed identification collapses.","supporting_citations":[{"cited_title":"Extracting point of interest and clas sifying environment for low sampling crowd sensing smartphone sens or data,","cited_arxiv_id":null,"evidence_quote":"Selects DBSCAN over other clustering algorithms for low-sampling crowd-sensed smartphone sensor data, motivating the paper's modified DBSCAN approach."},{"cited_title":"Wi-Fi Scanning,","cited_arxiv_id":null,"evidence_quote":"Sets the five-minute Wi-Fi scan interval as the battery-conscious standard in Android, which fixes the scan cadence and the minPts=4 parameter."},{"cited_title":"A density-based algorithm for discovering clusters in large spatial databa ses with noise","cited_arxiv_id":null,"evidence_quote":"Defines the original DBSCAN density-based clustering algorithm whose distance metric the paper replaces with cosine similarity."},{"cited_title":"Fast unfolding of communities in large networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Louvain modularity-optimization algorithm used to detect communities of POIs shared across users."}],"review_version":1}