{"id":"2675015c-e27c-4551-9906-14a691bbf9cb","arxiv_id":"2412.09292","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"The paper reports that a conditional GAN pre-trained on other homes and fine-tuned on a target home improves room-level BLE indoor localisation.","lead":"This paper trains a conditional GAN to generate synthetic Bluetooth signal data for indoor location tracking, then uses transfer learning so a model pre-trained on other homes can be fine-tuned for a new home. The authors report that this improves room-level localisation accuracy, especially in hard-to-cover areas like stairs and outdoor spaces.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The cross-home transfer claim rests on T-ConGAN-SPHERE, but no control shows the gain comes from RSSI content rather than from generic pretraining; the headline 12.2% figure comes from a same-protocol run that includes the target house in pre-training.","rationale":"The reader identifies the cross-protocol architecture adaptation as the weakest assumption and recommends CONDITIONAL. I agree that the adaptation is under-specified, but I see the more load-bearing gap as the absence of a matched control in the clean T-ConGAN-SPHERE experiment. The paper's strongest advertised number (12.2 percentage points) comes from T-ConGAN, where the target house is included in pre-training, so that number cannot support the cross-home sharing claim. The only clean cross-home evidence is T-ConGAN-SPHERE, and it does show improvement over baseline on all three target houses (B: 71.67 to 77.21; C: 68.96 to 79.42; D: 71.07 to 78.57), with non-overlapping standard deviations. However, compared with the no-transfer ConGAN, T-ConGAN-SPHERE is worse on house C (79.42 vs 79.60) and better on B and D, so the transfer benefit is inconsistent. More importantly, no control replaces the SPHERE pre-training data with structure-free data of the same shape, which is needed to rule out the possibility that any pretraining procedure or the architecture surgery itself provides the benefit. The paper does release code and repeats each experiment ten times, which is helpful, but the missing control directly targets the mechanism claimed in the abstract. The reader's concern about layer initialization is a specific instance of this broader issue: if the newly modified layers are randomly initialized and the middle layers dominate, we still do not know whether the shared RSSI content is what helps. A random-pretraining control would settle both the initialization question and the content question simultaneously. I therefore recommend keeping the CONDITIONAL verdict, with the added requirement that the clean cross-protocol experiment include a placebic pretraining control and that the same-protocol T-ConGAN be rerun with the target house excluded from pre-training.","tokens_in":11788,"tokens_out":7479,"duration_ms":82384,"concrete_test":"Run the T-ConGAN-SPHERE pipeline exactly as in Section 4.2.2 but replace the SPHERE pre-training windows with (a) Gaussian white noise of identical shape (9 APs × 20 timestamps) and (b) SPHERE windows whose time ordering and AP structure are randomly permuted, keeping the same number of pre-training steps, architecture modifications, and fine-tuning protocol. If the downstream macro-F1 scores on houses B, C, and D match the reported SPHERE-pretrained values (77.21, 79.42, 78.57), the RSSI content is not the active ingredient and the cross-home transfer claim is unsupported. Additionally, rerun T-ConGAN with the target house excluded from pre-training to test whether the headline 12.2-point gain persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of cross-home RSSI sharing is supported only by T-ConGAN-SPHERE (Section 4.2.2, Table 2), because the headline 12.2-point macro-F1 gain attributed to T-ConGAN is obtained by pre-training on houses B, C, and D and then fine-tuning on the same target house. For example, when house C is the target, house C's fingerprint data is also in the pre-training set, so the gain cannot distinguish transfer from simply exposing the model to the target house's own data during pre-training. The clean cross-protocol experiment pre-trains on three SPHERE homes (9 APs, unlabelled free-living data) and fine-tunes on houses B, C, D (11 APs), but no ablation or control verifies that the benefit comes from the RSSI-specific structure of the SPHERE data rather than from the pretraining procedure itself. The architecture modification (generator output channels 9 to 11, discriminator input channels 9 to 11, label embedding 3 to 10-11 classes) is substantial, and Section 4.2.2 does not state how the new layers are initialized or what fraction of the final performance is attributable to the retained weights. As a result, the evidence that 'shared RSSI information' rather than generic pretraining drives the gains is missing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a transfer-learning framework for RSSI-based indoor localisation, combining a Conditional GAN (ConGAN) for data augmentation with a transfer step (T-ConGAN). The idea is to pre-train a ConGAN on RSSI data from multiple houses—either houses from the same annotated dataset or unannotated SPHERE homes—and then fine-tune it on a target house to generate room-conditioned RSSI samples. The augmented data are used to train a Random Forest classifier for room-level localisation. Experiments on three residential houses report macro-F1 improvements of up to 12.2 percentage points and class-accuracy improvements of up to 51% for stairs and outdoor areas, and the paper claims to be the first to show that BLE RSSI data can be shared across homes with different experimental protocols.","tokens_in":12014,"tokens_out":5100,"duration_ms":51415,"significance":"If the cross-protocol transfer result is sound, the work has practical significance for reducing annotation effort in home health monitoring: pre-training on unannotated free-living data from other homes and fine-tuning on a small target dataset could improve localisation in minority classes such as stairs and outdoor spaces. The strengths of the paper are that it uses real residential data, repeats each experiment ten times with reported standard deviations, and makes the code available. However, the central claim is currently supported by an experiment that includes the target house in the pre-training pool, and the clean cross-protocol experiment lacks a control for generic pretraining. These issues are load-bearing for the claim that shared RSSI information, rather than pretraining itself, drives the improvement.","major_comments":[{"comment":"The same-protocol T-ConGAN experiment is confounded. The paper states that the pre-trained ConGAN uses combined fingerprint data from houses B, C, and D, with each house as a separate class, and is then transferred to a target house among B, C, and D. Thus, when house C is the target, house C's fingerprint data are also in the pre-training set. The 12.2 percentage-point gain for T-ConGAN on house C (81.16 vs. 68.96 in Table 2) cannot be attributed to cross-house transfer, because the model has seen the target house's own data during pre-training. Please retrain with the target house excluded from the pre-training pool, or explicitly reclassify this experiment as a warm-start result and remove it from the cross-house transfer claim.","section":"Section 4.2.2, Table 2"},{"comment":"The cross-protocol T-ConGAN-SPHERE experiment is the only non-circular support for the central claim, but it lacks a control for generic pretraining. The paper does not compare T-ConGAN-SPHERE with (i) the same architecture trained from scratch on the target house, or (ii) a model pre-trained on a shuffled or non-RSSI version of the SPHERE data. Without such a control, the gains over the plain ConGAN could arise from the pretraining procedure itself, such as better initialization or a learned label-embedding structure, rather than from shareable RSSI content. Please add an ablation that isolates the information content of the SPHERE RSSI signals.","section":"Section 5.2, Table 2"},{"comment":"The cross-protocol architecture adaptation is underspecified. When adapting from 9 SPHERE APs to 11 target-house APs, the paper modifies the generator output, discriminator input, and label embedding layers, but it does not state how these new layers are initialized, whether they are trained from scratch, or what fraction of the final model's weights is inherited from the SPHERE pre-training. This matters because the claimed mechanism is that pre-trained weights provide useful structure; if the new layers are randomly initialized and most of the retained weights are in lower layers, the result could be close to training from scratch. Please report the initialization scheme and an ablation that measures the contribution of the retained weights.","section":"Section 4.2.2, Figure 2"},{"comment":"The paper reports means and standard deviations over ten repetitions but provides no significance tests or effect sizes. Several comparisons in Table 2 are small relative to the reported deviations, for example T-ConGAN-SPHERE on house C (79.42 ± 0.57) versus SMOTE (77.91 ± 0.51), or T-ConGAN on house B (74.48 ± 0.79) versus Domain-Expert (74.72 ± 0.86). The claim that transfer-based methods 'consistently outperform' other augmentations would be strengthened by paired tests across the ten runs or by reporting confidence intervals.","section":"Table 2, Section 4.4"}],"minor_comments":[{"comment":"There is a stray space in 'T ransfer learning with ConGAN' in the section heading; also, the abstract contains 'required annotated data' where 'requires annotated data' would be grammatically correct.","section":"Section 4.2.2"},{"comment":"The line plots in Figure 3 are difficult to compare because the axes are not labeled and no RSSI scale is shown; adding axis labels and a consistent y-scale would improve interpretability.","section":"Figure 3"},{"comment":"The MiVo interpretation is not validated against the classification results: SMOTE has the lowest MiVo values but is among the best-performing methods in Table 2, so the statement that low MiVo suggests 'limited additional value' is not supported without a direct analysis linking MiVo to downstream classification performance.","section":"Section 5.1, Table 1"},{"comment":"Table 3 reports class-accuracy improvements but does not specify the baseline accuracies or standard deviations; please clarify whether these values are averaged over the ten runs and whether the improvements are stable across runs.","section":"Table 3"},{"comment":"House A is excluded from the experiments, but the paper does not explicitly discuss how this affects the generality of the transfer claims; please state this limitation more directly in the limitations section.","section":"Section 3.1, Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The same-protocol circularity is a serious issue that should be fixed before publication, and the cross-protocol result needs a pretraining control. I was not able to independently verify the 'first demonstration' claim against all prior work, so the editor may wish to ask the authors to substantiate the novelty claim with a more extensive related-work search. The paper is otherwise within scope and the core idea is promising, but the current evidence does not yet support the strong abstract-level claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The cross-protocol transfer idea is genuinely new and worth a look. Nobody has pre-trained a GAN on unannotated BLE RSSI from some homes and then fine-tuned it on a different home with a different AP count to improve room-level localisation. That is exactly what T-ConGAN-SPHERE does, and the result shows consistent macro-F1 gains over the no-augmentation baseline (77.21 vs 71.67 in house B, 79.42 vs 68.96 in C, 78.57 vs 71.07 in D). The code is public, evaluation uses held-out free-living data, and the repeated runs give standard deviations. That is real evidence and should be credited.\n\nThe soft spots are proportional. First, the headline 12.2% F1 gain (house C, 68.96 to 81.16) comes from the same-protocol T-ConGAN, which pre-trains on houses B, C, and D combined and then fine-tunes on C. The target house is in the pre-training pool, so that result cannot distinguish transfer from simply seeing the target data. The paper should not present it as the flagship claim. Second, the clean SPHERE cross-protocol experiment has no control. Fine-tuning from SPHERE pre-training might help because of a better initialization, not because the RSSI content carries useful structure. A control with random pre-training or pre-training on a different signal type would close that gap. Third, only three target homes are tested, with no significance tests; house C's gain is within overlapping standard deviations, so the claim of improvement there is fragile. Fourth, the 9-to-11 AP adaptation is under-specified: the paper does not say how the new generator output and discriminator input layers are initialized, making it unclear how much of the pre-trained model survives. Fifth, the 51% improvement is percentage points, not relative percent, and should be reported honestly. Finally, house A is excluded for stated reasons, but that carve-out should be flagged more prominently in the limitations.\n\nWho benefits: readers working on RSSI augmentation, transfer learning for wireless sensing, or smart-home health monitoring. The paper would benefit from a serious referee, but only after the same-protocol condition is rerun with the target removed, a random-initialization control is added, and significance is reported. I would not cite the current version for the transfer claim until that control exists.","headline":"The cross-protocol transfer idea is genuinely new and the SPHERE experiment is clean, but the headline 12.2% F1 gain comes from a confounded same-protocol run, so the paper needs a control and a rerun before the central claim holds.","tokens_in":12649,"tokens_out":2439,"would_cite":false,"duration_ms":25359,"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":"BLE RSSI from other homes can be transferred to improve indoor localisation in a new home, with macro F1 gains up to 12.2 points and 51-point accuracy gains in stairs and outdoors.","keywords":["indoor localisation","RSSI","BLE","transfer learning","conditional GAN","data augmentation","smart home","ConGAN"],"falsifier":"Run an ablation in which the pre-trained SPHERE weights are replaced with random initialisation at the same architecture, then fine-tune on house B/C/D fingerprint data; if the macro-F1 gain over single-house ConGAN disappears, the improvement is attributable to the transferred weights rather than to the architecture or fine-tuning procedure.","tokens_in":11503,"feed_emoji":"🏠","tokens_out":4866,"duration_ms":38358,"temperature":0.7,"pith_summary":"This paper tries to establish that Bluetooth Low Energy signal-strength readings (RSSI) collected in one set of homes carry reusable information for locating people in a different home, despite each house having a unique radio fingerprint. The authors propose T-ConGAN, which pre-trains a conditional GAN on RSSI from several source homes and then fine-tunes it on a small amount of labelled data from a target home to generate synthetic room-specific RSSI. They report that this transfer-based augmentation improves macro F1 for room-level localisation by up to 12.2 percentage points over training without augmentation, with the largest gains in hard-to-label areas like stairs and outdoor spaces. If correct, this would reduce the annotation burden for in-home health monitoring systems, since new deployments would not need hours of labelled walking tours.","feed_headline":"Shared BLE signals improve location tracking in new homes","feed_subtitle":"Pre-trained on other homes' RSSI, a conditional GAN lifts room-level F1 by up to 12 points.","key_machinery":"The load-bearing object is the transfer-learned Conditional Generative Adversarial Network (T-ConGAN): a ConGAN with 1D convolutional generator and discriminator, trained with Wasserstein loss and gradient penalty, pre-trained on source-home RSSI without room labels and then fine-tuned on target-home fingerprint data. The transfer step adapts three architectural pieces to the target: the label embedding layer (number of rooms), the generator's output layer (number of access points), and the discriminator's input layer. The pre-trained weights are intended to carry generic RSSI structure from the source homes, so that fine-tuning requires only small amounts of labelled target data. Evaluation couples the generator with a Random Forest classifier and the MiVo metric to compare synthetic data quality across augmentation methods.","core_discovery":"The central claim is that RSSI data can be shared across different homes, even when collected under different experimental protocols, and that the shared information improves indoor localisation performance. The authors demonstrate this with a conditional GAN (ConGAN) trained with Wasserstein loss and gradient penalty. They pre-train the ConGAN on RSSI from multiple source homes—either the same dataset's houses B, C, and D, or three SPHERE project homes with a different number of access points—then fine-tune it on a target house's fingerprint data to generate room-specific RSSI. Using a Random Forest classifier evaluated on held-out free-living data, they report macro F1 improvements of up to 12.2 percentage points and class accuracy improvements of up to 51 percentage points in stairs and outdoor areas, compared with no augmentation. They claim this is the first demonstration that BLE RSSI can be effectively shared across homes.","pith_inferences":["If the transfer effect holds broadly, one could pre-train a single large RSSI generator on many unlabelled homes and fine-tune it per residence, amortising the cost of annotation across an entire deployment fleet; the paper tests only three source homes, so scaling behaviour is an open question.","The reported 12.2-point macro-F1 gain is the difference between the best transfer method and the no-augmentation baseline in Table 2; a fairer comparison against a same-protocol transfer baseline (T-ConGAN) would isolate what cross-protocol data adds, which the paper does not fully separate.","A testable extension: use the same T-ConGAN framework with Wi-Fi or UWB RSSI, since the paper explicitly leaves cross-signal compatibility unexamined; the mechanism is signal-agnostic in principle."],"forward_implications":["New homes can deploy indoor localisation with substantially less labelled data, because a pre-trained generator supplies realistic room-specific RSSI after fine-tuning on minutes of fingerprint data.","Unannotated free-living RSSI datasets, such as the SPHERE homes used here, become useful training resources, extending the reach of augmentation beyond a single house.","Minority classes that are dangerous or hard to annotate—stairs, outdoor areas—gain the most, with reported accuracy improvements of 21–51 percentage points.","The method outperforms both classical augmentation (SMOTE, random oversampling, domain-expert noise) and single-house ConGAN augmentation, suggesting transfer learning is the active ingredient."],"supporting_citations":[{"why":"Provides the annotated residential RSSI dataset (houses A–D) used for target-house training and evaluation.","marker":"[5]"},{"why":"Supplies the unannotated free-living RSSI data from different homes and protocols used for cross-protocol pre-training.","marker":"[16]"},{"why":"Introduces conditional GANs, the base generator/discriminator conditioning mechanism used to generate room-specific RSSI.","marker":"[18]"},{"why":"Provides Wasserstein loss with gradient penalty, which stabilises ConGAN training.","marker":"[19]"},{"why":"Prior GAN-based RSSI augmentation for indoor localisation; the approach this paper extends and compares against.","marker":"[13]"},{"why":"Prior transfer-learning GAN (WGAN-GP) for WiFi localisation; the paper contrasts its within-building transfer with cross-house transfer.","marker":"[14]"}],"fun_headline_variants":["Sharing BLE signals across homes improves location accuracy","GAN-based transfer boosts indoor localisation by 12 percent","Cross-home RSSI transfer sharpens room-level positioning","T-ConGAN lifts F1 scores for in-home patient tracking","BLE transfer learning reduces need for annotated data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cross-protocol transfer result depends on the assumption that a GAN pre-trained on SPHERE homes with 9 access points can be adapted to target homes with 11 access points by resizing the generator output, discriminator input, and label embedding layers, while retaining enough pre-trained structure to help; the paper does not specify how the new layers are initialised or how much of the gain comes from the retained weights.","fun_headline_variants_meta":{"raw":{"variants":["Sharing BLE signals across homes improves location accuracy","GAN-based transfer boosts indoor localisation by 12 percent","Cross-home RSSI transfer sharpens room-level positioning","T-ConGAN lifts F1 scores for in-home patient tracking","BLE transfer learning reduces need for annotated data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000139,"raw_usage":{"total_tokens":1150,"prompt_tokens":931,"completion_tokens":219,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":142}},"tokens_in":547,"tokens_out":219,"duration_ms":3633,"temperature":1.0,"reasoning_tokens":142,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:05:57.715539+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an ablation in which the pre-trained SPHERE weights are replaced with random initialisation at the same architecture, then fine-tune on house B/C/D fingerprint data; if the macro-F1 gain over single-house ConGAN disappears, the improvement is attributable to the transferred weights rather than to the architecture or fine-tuning procedure.","supporting_citations":[{"cited_title":"Residential wearable rssi and accelerometer measurements with detailed location annotations,","cited_arxiv_id":null,"evidence_quote":"Provides the annotated residential RSSI dataset (houses A–D) used for target-house training and evaluation."},{"cited_title":"Sphere: A sensor platform for healthcare in a residential environment,","cited_arxiv_id":null,"evidence_quote":"Supplies the unannotated free-living RSSI data from different homes and protocols used for cross-protocol pre-training."},{"cited_title":"Improved training of wasserstein gans,","cited_arxiv_id":null,"evidence_quote":"Provides Wasserstein loss with gradient penalty, which stabilises ConGAN training."},{"cited_title":"Indoor localization using data augmentation via selective generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Prior GAN-based RSSI augmentation for indoor localisation; the approach this paper extends and compares against."},{"cited_title":"extendgan+: Transferable data augmentation framework using wgan-gp for data-driven indoor localisation model,","cited_arxiv_id":null,"evidence_quote":"Prior transfer-learning GAN (WGAN-GP) for WiFi localisation; the paper contrasts its within-building transfer with cross-house transfer."}],"review_version":1}