{"id":"589a2bb6-b60f-4022-bd6f-0e1e02c17679","arxiv_id":"2412.01979","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"FGATT, a fuzzy graph attention network paired with a Transformer encoder, is claimed to improve missing-data imputation on two SWaT wireless sensor datasets.","lead":"The paper proposes FGATT, a hybrid model that combines a fuzzy graph attention network with a Transformer encoder to fill missing values in wireless sensor data. The authors claim it outperforms four existing methods, especially when large portions of the data are missing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The forward pass for missing data is underspecified: Eq. (2) assumes complete node embeddings, yet the paper never defines embeddings for unobserved entries, so the reported robustness cannot be verified and may reflect mask artifacts rather than spatial dependencies.","rationale":"The reader's weakest assumption is that the fuzzy rough set connectivity score is unablated and may not be the source of gains. I agree with the need for an ablation, but I see a more fundamental gap: the paper does not specify what node embeddings are used when sensor readings are missing. Eq. (2) treats x_i^t as available, yet the task explicitly trains and tests with missing rates up to 80%. Without an explicit masking or input-imputation rule, the reported experiments are not reproducible and the robustness claim cannot be attributed to the described architecture. This reinforces the reader's REJECT verdict: the central empirical claim is unsupported as written. A targeted code inspection plus a static-graph ablation would settle whether the concern lands.","tokens_in":6651,"tokens_out":4647,"duration_ms":46176,"concrete_test":"Reproduce or inspect the code path for a test batch with 80% missing values: identify what value x_i^t takes for a missing sensor before Eq. (2) is evaluated. Then, keeping the architecture fixed, replace Eq. (2) with a static k-nearest-neighbor graph constructed from fully observed training data and compare MSE, MAE, and RMSE on SWaT.A7.22 at 50% and 80% missing rates. If the static graph matches or beats FGATT, the dynamic fuzzy graph is not responsible for the claimed robustness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical superiority at high missingness, but the model's forward pass under missingness is never specified. Section 3.1 defines connectivity scores as S^t(i,j) = alpha * R_B^{d_j}(x_i^t) + (1-alpha) * R_B^{d_i}(x_j^t) (Eq. 2), where x_i^t and x_j^t are node embeddings at time t. Section 4.2 states that training and testing use missing rates up to 80%, but no sentence describes what tensor is fed to the FGAT for unobserved entries: there is no mention of masking, zero-filling, a learned missing token, or a separate imputation layer. If unobserved values are replaced by a constant (e.g., zero after min-max normalization), the fuzzy lower and upper approximations in Eq. (1) will be dominated by the missingness indicator rather than by spatial signal, and the dynamic graph construction would learn connectivity from the mask pattern instead of from sensor relationships. That would invalidate the causal story that 'fuzzy rough sets-based dynamic graph captures meaningful spatial dependencies' (Section 4.3). In addition, the manuscript text contains no numeric results, error bars, or ablations; Figures 2 and 3 and the end of Section 4.3 are missing in the supplied text. Therefore the load-bearing condition—that FGATT's gains come from the described fuzzy graph mechanism under real missing data—is untested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FGATT, a framework for wireless sensor data imputation that combines a Fuzzy Graph Attention Network (FGAT) with a Transformer encoder. The FGAT component uses fuzzy rough sets to construct a dynamic graph from node embeddings, obviating the need for predefined spatial structure, while the Transformer models temporal dependencies. The authors claim that FGATT outperforms FFN, BGRU, Transformer, and TGCN baselines on SWaT sub-datasets at missing rates from 20% to 80%, and that the fuzzy rough set dynamic graph is responsible for the gains. The manuscript describes the proposed architecture, defines the connectivity score in Eq. (2), and outlines the experimental setup, but it does not provide the actual figures or numeric results, nor does it specify how missing values are represented in the model input.","tokens_in":6892,"tokens_out":3055,"duration_ms":29628,"significance":"If validated, the combination of self-adaptive fuzzy graph construction with a Transformer encoder would be a useful contribution to imputation in wireless sensor networks and IoT settings, where spatial topology is often unknown and missingness is high. The paper also has a plausible conceptual motivation: coupling fuzzy rough sets with graph attention to build dynamic connectivity can in principle capture uncertainty in sensor relationships. However, the significance depends entirely on the empirical claim of state-of-the-art accuracy and robustness, and that claim is not evidenced in the supplied text: Figures 2 and 3 are missing, no numeric MSE/MAE/RMSE values or error bars are reported, and the forward pass under missing data is underspecified. Because the paper's central contribution is empirical, these omissions are load-bearing rather than cosmetic.","major_comments":[{"comment":"The central claim—that FGATT outperforms all baselines in imputation accuracy and robustness—is unsupported because Figures 2 and 3 are absent and no numeric MSE, MAE, or RMSE values are reported in the text. Statements such as 'FGATT demonstrates the best average performance across all metrics' and 'outperforms baselines by a significant margin' cannot be verified. Please provide a complete results table (or actual figures) with the exact metric values for each dataset, each missing rate, and each baseline, along with standard deviations across multiple runs.","section":"§4.3, Figures 2 and 3"},{"comment":"The forward pass for missing data is not defined. Equations (1) and (2) compute connectivity scores from node embeddings x_i^t and x_j^t, but the manuscript never states how unobserved entries are represented in the input tensor—there is no mention of masking, zero-filling, a learned mask token, or a preliminary imputation layer. If missing values are replaced by a constant after min-max normalization, the fuzzy lower and upper approximations in Eq. (1) will be dominated by the missingness indicator rather than by sensor relationships, and the dynamic graph would learn connectivity from the mask pattern. Please specify the exact input representation for unobserved entries and justify that it preserves spatial signal.","section":"§3.1, Eqs. (1)-(2), and §4.2"},{"comment":"No ablation or sensitivity analysis isolates the contribution of the fuzzy rough set connectivity score to the reported gains. The abstract and Section 4.3 attribute improved robustness to 'fuzzy rough sets-based dynamic graph construction,' but the paper does not compare against a version of the model with a randomly generated graph, a distance-based graph, or a graph learned by a standard GAT layer. Please include ablations that vary the graph construction mechanism, the Transformer component, and hyperparameters such as alpha, top-K, and context window T to support the claimed causal role of the fuzzy graph.","section":"§4.3 and §3.2"},{"comment":"The experimental protocol is incomplete. The masking procedure (e.g., random vs. structured, MCAR/MAR, whether the same mask is used across methods), the number of independent runs, random seeds, and architecture details (number of layers, hidden sizes, dropout, learning rate) are not reported. Additionally, the claim in Section 4.3 that models 'overfit at the 50% missing rate' requires quantitative comparison of performance across missing rates. Without these details, the evaluation is not reproducible and the claimed trends cannot be assessed.","section":"§4.2"}],"minor_comments":[{"comment":"Because Eq. (2) and the pooling choice are taken verbatim from the authors' prior work [23] and [26], the novelty of FGATT beyond those papers should be stated explicitly in the contribution list.","section":"§1 and §2"},{"comment":"The notation R_B^{d_j}(x_i^t) relies on the definitions in Eq. (1), but the role of the decision attribute d_j in the connectivity score is not explained; please clarify how d_j is obtained for each node and time step.","section":"§3.1, Eq. (2)"},{"comment":"The sample counts for SWaT.A7.22 (3600) and SWaT.A7.29 (7201) are unusually small for SWaT, which normally contains much longer recordings; please confirm the selection procedure and explain these sub-dataset sizes.","section":"§4.1, Table 1"},{"comment":"The results section breaks off mid-sentence after 'the FGAT framework, which efficiently aggregates spatial and temporal'; the sentence should be completed and the discussion of Figure 3 finished.","section":"§4.3"},{"comment":"Reference [7] is a survey/catalog of transformer models rather than the original transformer paper; please cite the primary source for the Transformer architecture.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript as supplied appears to be an incomplete version of a conference submission—Figures 2 and 3 are missing and Section 4.3 is truncated. This alone explains the lack of numeric results, but even with those figures provided, the authors would need to define the missing-data input representation and add ablations. The reference list contains several self-citations from the same group, and the core equations are imported from two of those papers; while not disqualifying, the novelty boundary should be verified carefully in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take: FGATT is not ready for review. The idea—pairing the authors' prior fuzzy graph attention construction with a Transformer encoder for wireless imputation—is reasonable and incremental. The SWaT evaluation with varying missing rates is a sensible testbed. But the manuscript provided here contains no numeric results, no error bars, no code, and Figures 2 and 3 are missing; Section 4.3 ends mid-sentence. The abstract's claim that FGATT 'outperforms state-of-the-art methods' is therefore unsupported. That alone justifies rejection in this form.\n\nThe more serious problem is conceptual, not just presentational. Section 3.1 defines connectivity scores over node embeddings x_i^t and x_j^t, but nowhere does the paper say what those embeddings are when the corresponding sensor reading is missing. Training and testing use missing rates up to 80%, yet there is no mention of masking, zero-filling, a learned missing token, or a separate imputation layer. If unobserved entries are replaced by a constant, the fuzzy lower and upper approximations in Eq. (1) will be dominated by the pattern of missingness rather than by genuine spatial dependencies, and the reported 'dynamic graph' would be learning from the mask. The paper even attributes its gains to 'fuzzy rough sets-based dynamic graph construction capturing meaningful spatial dependencies' (Section 4.3), so this is load-bearing. The forward pass has to be specified before any claim of robustness can be evaluated.\n\nWhat the paper does well: it is honest about borrowing Eq. (2) and the FGAT layer from the authors' own prior work [23] and the pooling choice from [26], and the related work is standard. The idea of applying fuzzy rough sets to dynamic graph construction for imputation is not silly, and the high-missingness regime is a real problem. But novelty is incremental, and the reliance on self-cited components means the new combination is the only real contribution.\n\nMy verdict: desk reject. The manuscript is not complete enough to send to referees; they would spend their time reconstructing missing results. I would ask the authors for a version with actual numbers, explicit masking handling, and an ablation separating the contributions of the fuzzy graph, the GAT, and the transformer. If those come back, the paper might be worth a look at a workshop or a venue with lower bars.","headline":"Incomplete manuscript: the central empirical claim is unverifiable as written, and the forward pass for missing data is never defined.","tokens_in":7486,"tokens_out":2282,"would_cite":false,"duration_ms":21236,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FGATT combines fuzzy graph attention with a Transformer encoder and reports the best imputation accuracy on SWaT water-treatment data across missing rates from 20% to 80%.","keywords":["data imputation","fuzzy graph attention networks","Transformer encoder","wireless sensor networks","dynamic graph construction","fuzzy rough sets","spatio-temporal modeling","missing data"],"falsifier":"Keep FGATT's architecture fixed but replace the connectivity score in Eq. (2) with random edge weights or with a static distance-based graph on the same SWaT datasets; if the MSE, MAE, and RMSE curves at 20% to 80% missingness change only slightly, the fuzzy-rough-set graph is not the source of the reported advantage.","tokens_in":6374,"feed_emoji":"📡","tokens_out":6242,"duration_ms":276331,"temperature":0.7,"pith_summary":"This paper proposes FGATT, a two-stage neural architecture for filling missing wireless sensor readings. The first stage adaptively builds a graph among sensors, using fuzzy rough sets to score how strongly one sensor's reading relates to another, so no predefined spatial map is needed. The second stage passes graph-refined representations through a Transformer encoder to capture temporal dependencies. The authors report the lowest mean squared error, mean absolute error, and root mean squared error on two SWaT benchmark subsets at most tested missing rates from 20% to 80%, with the gap over baselines widening when more than half the data is absent. A sympathetic reader would take the contribution to be a practical, map-free spatial-temporal imputation module for wireless and IoT settings.","feed_headline":"Fuzzy graph attention plus Transformer wins wireless imputation tests","feed_subtitle":"On two SWaT sensor subsets, the hybrid model keeps errors lowest even above 50 percent missing data.","key_machinery":"The load-bearing mechanism is a self-adaptive graph constructed from fuzzy rough-set connectivity scores. For each pair of nodes, a score is computed as a weighted combination of fuzzy lower- and upper-approximation memberships, averaged over a 16-timestep context window, and only the top-$K$ edges are retained. A GAT layer aggregates neighbor information over this dynamically built graph, and a Transformer encoder's self-attention then captures temporal dependencies. This design replaces the static, predefined graphs used by earlier spatio-temporal models with a connectivity structure learned directly from the data.","core_discovery":"The central claim is that FGATT, the hybrid of fuzzy rough-set dynamic graph construction, GAT spatial aggregation, and a Transformer encoder, outperforms the compared baselines in imputation accuracy and robustness, particularly at high missingness. On SWaT.A7.22, the paper reports that FGATT demonstrates the best average performance across all metrics, and on SWaT.A7.29 it consistently achieves the best performance under all metrics. The authors attribute this advantage to the dynamic graph's ability to capture spatial dependencies without predefined spatial information, while the Transformer encoder models temporal patterns. If correct, this would make FGATT a useful drop-in imputation tool for wireless sensor networks where sensor maps are unavailable or unreliable.","pith_inferences":["A natural extension the authors do not test is streaming imputation: because connectivity scores are computed per timestep, the graph can in principle be updated online as new sensor readings arrive, provided the pooling window stays short.","The same adaptive graph construction could transfer to non-geospatial multivariate time series such as finance or health monitoring, where 'spatial' dependence is functional rather than physical; the paper's experiments do not demonstrate this transfer.","The choice of mean pooling is borrowed from sentiment-analysis pooling comparisons; a drop-in ablation replacing it with max or weighted-sum pooling would reveal whether the aggregation strategy affects imputation quality.","If the reported gains persist when the fuzzy-rough connectivity score is replaced by a simpler learned distance metric, then the fuzzy-rough component is not the active ingredient; the paper reports no such control, making that a testable implication of its robustness claim."],"forward_implications":["At test missing rates below the 50% training rate, FGATT keeps errors stable, so a model trained once at high missingness can serve deployments with less missing data.","Above 50% missingness, FGATT degrades more gracefully than the baselines, and TGCN shows the steepest drop, suggesting static-graph spatial models are the weakest in sparse regimes.","Because the graph is built from data rather than sensor coordinates, the method applies to datasets where spatial metadata is missing or untrusted.","The encoder-only Transformer design is presented as a balance between computational cost and information richness, so the architecture is intended to scale to longer sensor sequences.","The paper's evaluation protocol on the two SWaT subsets provides a reusable benchmark for future imputation work: normalization, a 70/10/20 split, context length 16, and missing-rate sweeps from 20% to 80%."],"supporting_citations":[{"why":"Supplies the FGAT layer and the connectivity-score formula in Eq. (2), including top-K edge retention and self-loop removal, which are the paper's core spatial modeling components.","marker":"[23]"},{"why":"Provides the fuzzy lower and upper approximation definitions that the connectivity score is built on.","marker":"[24]"},{"why":"Supplies the kernel-based fuzzy rough relation and an earlier application of weighted fuzzy rough sets, grounding Eq. (1).","marker":"[25]"},{"why":"Justifies the choice of mean pooling when aggregating connectivity scores over the temporal context window.","marker":"[26]"},{"why":"Provides the SWaT water-treatment dataset used for both experimental evaluations.","marker":"[27]"},{"why":"The BGRU baseline, representing a bidirectional temporal recurrent model that FGATT is compared against.","marker":"[28]"},{"why":"The TGCN baseline, a spatial-temporal graph network that FGATT must beat, especially at high missing rates.","marker":"[29]"},{"why":"The Transformer baseline, the temporal model whose encoder structure FGATT extends.","marker":"[7]"}],"fun_headline_variants":["FGATT hybrid beats baselines for wireless imputation","Robust imputation: fuzzy graph attention and transformer team up","No sensor map? FGATT learns graph and imputes wirelessly","Fuzzy graph + transformer: best imputation under heavy missingness","FGATT: dynamic graphs and transformer for resilient imputation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the dynamic graph built from the fuzzy rough-set connectivity score in Eq. (2) captures genuine spatial relationships among the sensors, so that the graph module, rather than the Transformer or the GAT attention itself, is what drives the reported accuracy gains.","fun_headline_variants_meta":{"raw":{"variants":["FGATT hybrid beats baselines for wireless imputation","Robust imputation: fuzzy graph attention and transformer team up","No sensor map? FGATT learns graph and imputes wirelessly","Fuzzy graph + transformer: best imputation under heavy missingness","FGATT: dynamic graphs and transformer for resilient imputation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000988,"raw_usage":{"total_tokens":4142,"prompt_tokens":853,"completion_tokens":3289,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":3203}},"tokens_in":469,"tokens_out":3289,"duration_ms":24587,"temperature":1.0,"reasoning_tokens":3203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:57:41.245177+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep FGATT's architecture fixed but replace the connectivity score in Eq. (2) with random edge weights or with a static distance-based graph on the same SWaT datasets; if the MSE, MAE, and RMSE curves at 20% to 80% missingness change only slightly, the fuzzy-rough-set graph is not the source of the reported advantage.","supporting_citations":[{"cited_title":"Weighted fuzzy rough sets-based tri- training and its application to medical diagnosis","cited_arxiv_id":null,"evidence_quote":"Supplies the kernel-based fuzzy rough relation and an earlier application of weighted fuzzy rough sets, grounding Eq. (1)."},{"cited_title":"SWaT: A water treatment testbed for research and training on ICS security","cited_arxiv_id":null,"evidence_quote":"Provides the SWaT water-treatment dataset used for both experimental evaluations."},{"cited_title":"Time-series production forecasting method based on the integration of Bidirectional Gated Recurrent Unit (Bi-GRU) network and Sparrow Search Algorithm (SSA)","cited_arxiv_id":null,"evidence_quote":"The BGRU baseline, representing a bidirectional temporal recurrent model that FGATT is compared against."},{"cited_title":"T-GCN: A temporal graph convolutional network for traffic prediction","cited_arxiv_id":null,"evidence_quote":"The TGCN baseline, a spatial-temporal graph network that FGATT must beat, especially at high missing rates."}],"review_version":1}