{"id":"ff6b7f21-c14d-48be-a7e9-60515bcccf87","arxiv_id":"2501.18221","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A network-weighted functional regression with conformal prediction bands is presented for functional data on graphs; coverage guarantees hold only for the supremum-norm score, not the proposed L2 score.","lead":"The paper proposes a regression model for functional data measured at the nodes of a network, where each node's prediction is influenced by its neighbors through graph-based weights, together with a conformal prediction method for uncertainty bands. It is aimed at settings like environmental sensor networks, where signals at nearby sensors are dependent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised finite-sample coverage guarantee is invalid for the proposed D2 pointwise band: the conformal set for an L2 score is an L2 ball, not the reported ±k_S S(t) band, and Table 4 itself shows 80–86% global coverage at α=0.05.","rationale":"The estimation part of NWFR, summarized in Theorem 3.1, is a direct weighted least-squares extension of GWFR and is not the basis for rejection. The decisive issue is the conformal coverage claim. Section 4.1 uses an L2-based nonconformity score to select k_S but reports a pointwise band. Split conformal prediction constructs the prediction set as the sublevel set of the nonconformity score, so for D2 the set is an L2 ball. The paper offers no argument that the pointwise band equals, contains, or is calibrated to this ball; in fact the two sets are not nested in general. Hence the coverage guarantee for the score cannot be transferred to the reported band. Table 4's global coverage of 80–86% for D2 at α=0.05 is direct empirical evidence: if the guarantee were operative for the reported band, CovG would be at least 95% (up to Monte Carlo error). The D∞ variant is different because its sublevel set is the pointwise band, and its roughly 95–96% coverage in the same table is consistent with that. This makes the work salvageable if the coverage claim is restricted to D∞ or replaced by the conformal L2 ball, but the abstract's unconditional 'guaranteed coverage' for the proposed procedure is not supported. I agree with the reader's REJECT verdict, so no change is needed; the D2-versus-pointwise-band mismatch is the same load-bearing concern the reader identified.","tokens_in":16678,"tokens_out":6098,"duration_ms":58121,"concrete_test":"Run a clean i.i.d. version of the Section 4.1 split-conformal procedure (exchangeable nodes, n_cal = 200, n_test = 1000), computing k_S from D2 scores. For the same fitted model, compare three sets: (i) the reported pointwise band Yhat ± k_S S; (ii) the conformal L2 ball {Y : ∫((Yhat-Y)/S)^2 ≤ k_S^2}; (iii) the D∞ band Yhat ± k_∞ S. Report CovG at α=0.05 for each. If (ii) and (iii) are near 95% while (i) is near 80–86%, the pointwise band is not the conformal set for D2 and the guarantee cannot hold for it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is in Section 4.1. The paper defines the nonconformity score D2(Yhat,Y) = sqrt(∫_T ((Yhat(t)-Y(t))/S(t))^2 dt), takes k_S as its (1-α)-quantile, but then constructs the prediction set as the pointwise band {Y : Yhat(t)-k_S S(t) ≤ Y(t) ≤ Yhat(t)+k_S S(t) for all t}. For h=2, the conformal set {Y : D2(Yhat,Y) ≤ k_S} is an L2 ball around Yhat, not a pointwise band. Membership in the reported band is neither necessary nor sufficient for membership in that ball: a curve can exit the band at one time point yet have small L2 score, and a curve can stay inside the pointwise envelope while accumulating enough L2 mass to have D2 > k_S. Therefore, even if the network-stratified exchangeability assumption in Section 4.1 is granted, the finite-sample coverage guarantee applies to the L2 ball, not to the band the paper actually reports and evaluates. The paper's own Table 4 is the empirical manifestation: for D2 at α=0.05, global coverage CovG ranges from 80.5% to 85.8%, far below 95%, whereas D∞, whose conformal set is exactly the pointwise band, gives about 95–96%. This is an internal inconsistency in the central claim, not merely a departure from current consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Network-Weighted Functional Regression (NWFR), an extension of geographically weighted functional regression to data observed on a graph. The model estimates node-specific functional coefficient surfaces via weighted least squares, with weights determined by a Gaussian kernel of the geodesic distance on the network. The paper further proposes a split-conformal prediction procedure for functional responses at new vertices, using either an L2-based nonconformity score D2 or a sup-norm score D∞, and claims distribution-free finite-sample prediction bands with guaranteed coverage. The method is evaluated on twelve simulated scenarios with network-structured functional data and on the Intel Berkeley indoor sensor dataset, comparing NWFR against classical functional regression and geographically weighted functional regression. The paper reports improvements in point prediction accuracy and discusses a coverage-efficiency trade-off between D2 and D∞.","tokens_in":16984,"tokens_out":4321,"duration_ms":38046,"significance":"If the stated guarantees held, NWFR would be a useful addition to the literature on second-generation functional data, combining network-aware functional regression with distribution-free uncertainty quantification. The algebraic derivation of the coefficient estimator in Theorem 3.1 (Eq. 7 from Eq. 6) is correct for a fixed weight matrix, and the simulation study is extensive (12 scenarios, 100 replicates). The paper also usefully reports several evaluation metrics (CovG, CovL, ABW, interval score) and is explicit about the validity-efficiency trade-off. However, the central claim of guaranteed coverage for the proposed D2-based pointwise band is invalid, and the paper's own Table 4 empirically contradicts it. The D∞ variant does achieve approximately nominal coverage, but that variant is a direct extension of existing functional conformal methods and is not the paper's proposed nonconformity score.","major_comments":[{"comment":"The prediction set defined as C(X_{ν_j}) = {y(t) : Ŷ_{ν_j}(t) − k_S S(t) ≤ y(t) ≤ Ŷ_{ν_j}(t) + k_S S(t) for all t} is not the conformal set associated with the D2 nonconformity score. For D2(Ŷ, Y) = sqrt(∫_T ((Ŷ(t) − Y(t))/S(t))^2 dt), the conformal set is {Y : D2(Ŷ, Y) ≤ k_S}, which is an L2 ball in the metric weighted by 1/S(t), not a pointwise band. Membership in the pointwise band is neither necessary nor sufficient for membership in that ball: a curve can leave the band at a single time point yet have small L2 score, and a curve can remain inside the pointwise envelope while accumulating enough L2 mass to have D2 > k_S. Consequently, even if the exchangeability assumption in Section 4.1 is granted, the finite-sample coverage guarantee applies to the L2 ball, not to the band the paper constructs and evaluates. Table 4 confirms this: for D2 at α = 0.05, global coverage CovG ranges from 80.5% to 85.8% across the twelve scenarios, far below 95%, whereas D∞, whose conformal set is exactly the pointwise band, gives approximately 95–96% coverage. This is an internal inconsistency in the paper's central claim.","section":"Section 4.1, Step 5"},{"comment":"The validity of the entire conformal procedure rests on the assertion that there exists a partition G_1,...,G_K of the vertex set such that within each group the pairs (X_{ν_i}, Y_{ν_i}) are exchangeable, and that drawing at most one vertex from each group into the calibration set 'recovers the exchangeability needed for valid conformal prediction.' This is stated without proof or a precise description of the sampling scheme. Moreover, the Louvain partition used in Step 1 is estimated from the same data, which can itself induce dependence among the sampled units; no result is given showing that the estimated partition satisfies the required exchangeability condition with high probability or in expectation. The citation to Lunde et al. [27] provides intuition about network-aware sampling but does not establish the assertion for the proposed split-conformal algorithm with an estimated community structure. Since the finite-sample coverage guarantee in Step 5 depends on this assumption, it is a load-bearing gap.","section":"Section 4.1, paragraph on partition"}],"minor_comments":[{"comment":"There are several typos: 'asses' should be 'assess', 'uncertainity' should be 'uncertainty', and 'on on appropriately modelling' should be 'on appropriately modelling'.","section":"Abstract and Introduction"},{"comment":"The description of the Connectivity Between Communities parameter is contradictory: 'Communities inter link probability is 0.2 (CaseHigh), respectively the0.5 (Case High)' should presumably read '0.2 (Case Low) and 0.5 (Case High)'.","section":"Section 5, parameter CBC"},{"comment":"The text refers to 'Table 6' for the model goodness-of-fit metrics, but the table is labeled Table 5; the p-values table is Table 6, not Table 7 as stated in the text.","section":"Section 6, table cross-references"},{"comment":"The definition of S(t) contains a duplicated phrase: 'Formally, the modulation function can be defined as: as:' should be 'Formally, the modulation function can be defined as:'.","section":"Section 4.1, modulation function"},{"comment":"The sup-norm score D∞ is used in the tables but is never formally defined; the paper should state explicitly that D∞ is the limit h→∞ of D_h, i.e., sup_{t∈T} |Ŷ(t) − Y(t)| / S(t).","section":"Section 4.1, nonconformity scores"}],"recommendation":"reject","confidential_remarks":"The central flaw is not a presentation issue: the paper's headline contribution—D2-based pointwise prediction bands with guaranteed coverage—is mathematically inconsistent, because the conformal set for an L2 score is not the reported band. The paper's own Table 4 shows the empirical consequence. While the D∞ variant achieves nominal coverage, that variant is a minor extension of existing functional conformal methods and does not salvage the paper's stated novelty. The exchangeability assumption for the estimated Louvain partition also lacks any supporting argument. I recommend rejection, as the required changes go beyond a revision of presentation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the NWFR estimation part is a clean extension of GWFR to graph geodesics, and the simulations are thorough, but the headline coverage guarantee doesn't survive contact with the actual nonconformity score. For the D2 score the conformal set is an L2 ball, not the pointwise band Yhat ± kS S(t). The paper's own Table 4 shows CovG around 80–86% for D2 at α=0.05, while D∞ gives ~95–96%. That's not a subtle numerical gap; it's the wrong set being reported as the conformal band.\n\nWhat's good: Eq. (7) is a standard weighted least squares solution and is correctly derived. The paper demonstrates large gains in R² over classical functional regression across 12 simulated scenarios, and the Intel lab data application is a reasonable real-world test. The local coverage measure CovL is a useful addition. Citations to prior conformal functional work are appropriate, including their own [12].\n\nSoft spots: the exchangeability assumption behind the community-stratified calibration is asserted rather than proven; drawing at most one vertex per community doesn't automatically recover exchangeability, especially when the partition is estimated from the same data (Louvain). The simulations never compare against GWFR, which would be the natural baseline to show that network-geodesic weights beat spatial weights. The real-data section doesn't clearly describe how the conformal split was done.\n\nBottom line: the paper has a load-bearing flaw in its central claim, but it is fixable—use D∞ for pointwise bands, or report the L2 ball as the conformal set for D2. The estimation methodology is a legitimate contribution and the paper is worth a serious referee, likely leading to major revision. I'd bring it to reading group as a cautionary tale about matching nonconformity scores to the reported prediction set.","headline":"The coverage guarantee advertised for the D2 pointwise band is invalid—the conformal set is an L2 ball—and the paper's own simulations show it; the D∞ version works, and the estimation part is sound.","tokens_in":17571,"tokens_out":3009,"would_cite":false,"duration_ms":27811,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62R10","62M30","62G15","62F40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Network-Weighted Functional Regression (NWFR) weights each node's functional regression by graph geodesic distances, and its conformal wrapper claims distribution-free, finite-sample prediction bands for the whole response curve at a new…","keywords":["functional data analysis","network data","conformal prediction","functional regression","geographically weighted regression","prediction intervals","community detection","sensor networks"],"falsifier":"Simulate networks where within-community pairs are generated with clear hub-dependent variances, use the paper's Louvain split-conformal procedure with the $D_2$ score, and record the empirical fraction of complete curves contained in $\\hat{Y}_{\\nu_j}(t) \\pm k_S S(t)$ on a large test set; if that fraction is substantially below $1-\\alpha$ while the fraction of curves inside the $L^2$ ball of radius $k_S$ is at least $1-\\alpha$, the pointwise band is not the conformal set and the stated guarantee fails.","tokens_in":16367,"feed_emoji":"🌐","tokens_out":10461,"duration_ms":86790,"temperature":0.7,"pith_summary":"Functional data that live on a network—sensor readings, brain signals, social measurements—violate the independence assumption of classical functional regression because nearby nodes tend to behave alike. This paper introduces the Network-Weighted Functional Regression (NWFR) model, which replaces the global least-squares fit with a node-wise weighted fit: the influence of another node on the fit at a given node is a Gaussian kernel of the shortest-path distance between them on the graph. The authors report that on simulated stochastic-block networks and on indoor micro-climate sensor data, NWFR raises explained variability from roughly 20% for the unweighted model to above 86% in nearly all scenarios, and it outperforms the spatial GWFR model. To quantify uncertainty, the paper develops a split-conformal prediction procedure in which the vertices are first stratified into communities (via Louvain) to recover approximate exchangeability, and a calibration-set quantile $k_S$ is used to form a prediction band $\\hat{Y}_{\\nu_j}(t) \\pm k_S S(t)$ claimed to contain the full response curve at a new vertex with probability at least $1-\\alpha$. The purpose of the paper is to establish that explicitly modeling the network structure yields both better point forecasts and honest, distribution-free prediction intervals for functional responses on networks.","feed_headline":"Network functional regression yields coverage-guaranteed bands","feed_subtitle":"Node fits are weighted by graph distance; conformal prediction supplies claimed coverage guarantees.","key_machinery":"The central object is the node-specific diagonal weight matrix $W_{\\nu_i}$, whose diagonal entries are $w(i,j) = \\exp\\left\\{-\\frac{1}{2}(d_{ij}/\\theta)^2\\right\\}$, with $d_{ij}$ the shortest-path (geodesic) distance between vertices $\\nu_i$ and $\\nu_j$ on the weighted graph and $\\theta$ a bandwidth chosen by cross-validation. Given B-spline bases for the covariates and the response, the fitted coefficient block at node $\\nu_i$ has the closed form $B_{\\nu_i} = (X^T W_{\\nu_i} X)^{-1} X^T W_{\\nu_i} Y$, a network-weighted version of the functional concurrent regression estimate. The conformal machinery has three parts: a community partition of the vertices (via Louvain) from which at most one vertex per community is drawn into the calibration set; a modulation function $S(t)$, the pointwise root-mean-square calibration error; and the nonconformity scores $D_h = \\left(\\int_T \\left( \\frac{\\hat{Y}_{\\nu_i}(t)-Y_{\\nu_i}(t)}{S(t)} \\right)^h dt\\right)^{1/h}$, with $h=2$ and $h=\\infty$ compared. A permutation test on the integrated variance of the coefficient functions across nodes tests whether the network structure has a significant effect on each covariate's coefficient surface.","core_discovery":"The paper claims that functional regression on network-indexed data can be made both more accurate and more reliably quantified by (i) replacing the global least-squares fit with a node-wise weighted fit in which the weight between two nodes is a Gaussian kernel of their geodesic distance on the graph, and (ii) surrounding the fitted curves with conformal prediction bands built from an exchangeability-stabilizing partition of the vertices into communities. The authors state that for a new vertex $\\nu_j$, with probability at least $1-\\alpha$, the whole curve $Y_{\\nu_j}$ lies inside the band $\\hat{Y}_{\\nu_j}(t) \\pm k_S S(t)$, where $k_S$ is the $(1-\\alpha)$-quantile of the nonconformity scores and $S(t)$ is a data-driven modulation function. Two scores are considered: the $L^2$-based $D_2$ score, which averages squared errors over the domain, and the supremum-based $D_\\infty$ score, which focuses on the worst point. In their simulations and in the Intel sensor micro-climate case study, the conformal NWFR bands typically exceed 95% global coverage when $D_\\infty$ is used, while $D_2$ gives narrower bands with lower coverage, and the paper interprets this as the standard validity-efficiency trade-off.","pith_inferences":["The paper's coverage guarantee is stated for the pointwise band $\\hat{Y}_{\\nu_j}(t) \\pm k_S S(t)$ together with the $L^2$ score $D_2$, but the conformal set associated with an $L^2$ norm is an $L^2$ ball, not a pointwise envelope; a separate argument would be needed to transfer the coverage guarantee to the reported band.","The Louvain partition is estimated from the same data that are later used for calibration, so the 'at most one vertex per community' rule conditions on a data-dependent grouping; treating the partition as fixed rather than estimated could understate the coverage risk.","A natural stress test is to study sensitivity to the kernel bandwidth $\\theta$: network regions with heterogeneous edge densities may require locally varying bandwidths, and the single global $\\theta$ could limit the model's flexibility.","The method assumes the new vertex's graph connections (and hence its geodesic distances) are known at prediction time; extending the framework to nodes that join the graph later, or to entirely unseen components, would require estimating distances or a different weight construction."],"forward_implications":["For network-indexed functional data, NWFR supplies node-specific coefficient surfaces and predictions that explicitly encode graph proximity; a new node's response curve can be predicted from its covariates once its geodesic distances to the training nodes are known.","If the conformal coverage claim holds, practitioners get finite-sample marginal coverage without distributional assumptions, so uncertainty bands remain meaningful for non-Gaussian error processes on networks.","The comparison of $D_2$ and $D_\\infty$ scores makes the validity-efficiency trade-off explicit: supremum-based scores give higher global coverage with wider bands, while $L^2$-based scores give narrower bands at the cost of coverage.","The permutation test on coefficient variance provides a direct check of whether the network topology materially changes the regression coefficients for a given covariate.","The bandwidth $\\theta$ of the Gaussian geodesic kernel controls the neighborhood scale; cross-validation selects it, and the model reduces to the classical concurrent functional regression when all weights are equal."],"supporting_citations":[{"why":"Supplies the geographically weighted functional regression model and the Gaussian distance kernel that NWFR extends to network settings.","marker":"[40]"},{"why":"Supplies the network-aware sampling and stratification technique used to restore exchangeability in the conformal procedure.","marker":"[27]"},{"why":"Supplies the functional split-conformal prediction framework and the supremum-based nonconformity score that the paper adapts.","marker":"[13]"},{"why":"Justifies applying conformal prediction when exchangeability is violated, which underpins the coverage claim on networks.","marker":"[5]"},{"why":"The Louvain algorithm used to detect communities for network stratification.","marker":"[7]"},{"why":"Provides the local spatial conformal prediction approach and the principle of leveraging empirical variability for valid bands.","marker":"[28]"},{"why":"Provides the distribution-free predictive inference framework underlying the quantile-based band construction.","marker":"[25]"},{"why":"Provides the coverage and interval-score evaluation metrics (CovG, CovL, ABW, interval score) used in the study.","marker":"[12]"}],"fun_headline_variants":["Network functional regression gains guaranteed coverage bands","Graph-weighted functional fits with valid conformal intervals","D_inf yields coverage, D2 yields narrow bands","Coverage-guaranteed prediction intervals for network curves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that grouping the network's nodes into communities and taking at most one node per community for calibration makes the observations interchangeable in the way conformal prediction requires, and that the reported pointwise band is exactly the prediction set; neither is proved, and the grouping is estimated from the same data.","fun_headline_variants_meta":{"raw":{"variants":["Network functional regression gains guaranteed coverage bands","Graph-weighted functional fits with valid conformal intervals","D_inf yields coverage, D2 yields narrow bands","Coverage-guaranteed prediction intervals for network curves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000327,"raw_usage":{"total_tokens":1805,"prompt_tokens":895,"completion_tokens":910,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":850}},"tokens_in":511,"tokens_out":910,"duration_ms":9378,"temperature":1.0,"reasoning_tokens":850,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:16:14.126264+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate networks where within-community pairs are generated with clear hub-dependent variances, use the paper's Louvain split-conformal procedure with the $D_2$ score, and record the empirical fraction of complete curves contained in $\\hat{Y}_{\\nu_j}(t) \\pm k_S S(t)$ on a large test set; if that fraction is substantially below $1-\\alpha$ while the fraction of curves inside the $L^2$ ball of radius $k_S$ is at least $1-\\alpha$, the pointwise band is not the conformal set and the stated guarantee fails.","supporting_citations":[{"cited_title":"Yamanishi and Y","cited_arxiv_id":null,"evidence_quote":"Supplies the geographically weighted functional regression model and the Gaussian distance kernel that NWFR extends to network settings."},{"cited_title":"Diquigiovanni, M","cited_arxiv_id":null,"evidence_quote":"Supplies the functional split-conformal prediction framework and the supremum-based nonconformity score that the paper adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the local spatial conformal prediction approach and the principle of leveraging empirical variability for valid bands."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the distribution-free predictive inference framework underlying the quantile-based band construction."},{"cited_title":"Diana, E","cited_arxiv_id":null,"evidence_quote":"Provides the coverage and interval-score evaluation metrics (CovG, CovL, ABW, interval score) used in the study."}],"review_version":1}