{"id":"ae2913d6-2acc-4838-85af-61996f581ee6","arxiv_id":"2501.10151","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"TDAR combines topology-based attribute pre-filling, dynamic node weighting, and homophily regularizers to improve attribute recovery and downstream performance on attribute-missing graphs.","lead":"This paper introduces TDAR, a graph-learning method that fills in missing node attributes by spreading known attribute values across graph connections. It reports small but consistent gains over prior methods on standard citation and product graphs, with code released on GitHub.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TDAR's central robustness claim assumes homophily, yet only homophilous citation/product graphs are tested; on heterophilous SIoT-style graphs its propagation and contrastive losses could reverse the reported gains.","rationale":"The paper's own formulations make the homophily assumption explicit: TAAP is derived from Dirichlet energy minimization, NHS sums logistic similarity over edges, and NLSC suppresses similarity over non-edges. The reported experiments are internally consistent, the code is provided, and the gains on the tested datasets look plausible, so there is no reason to reject the empirical claim for homophilous graphs. However, the central narrative is framed around SIoT, where edges do not always imply attribute or label similarity. The most load-bearing gap is that the method's inductive bias is never tested against heterophilous or SIoT-like graphs, so the claimed general robustness is unestablished. This matches the reader's weakest assumption and supports the CONDITIONAL verdict; no change to the reader's verdict is required.","tokens_in":24512,"tokens_out":7069,"duration_ms":76686,"concrete_test":"Run TDAR under the paper's masking protocol (40/10/50 split, hyperparameters from §V-C) on three standard heterophilous benchmarks—for example Chameleon, Squirrel, and Roman-empire—and compare against the strongest baselines (SVGA, MATE, AIAE) using the same Recall@K, nDCG, ACC, and NMI metrics. If TDAR's reconstruction or clustering results fall below these baselines on any heterophilous graph, the homophily assumption is load-bearing and the general SIoT robustness claim must be narrowed or revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The three core mechanisms of TDAR all encode a homophily/smoothness prior. TAAP (Eq. 8) propagates known features along edges as a Dirichlet-energy-minimizing pre-fill, NHS (Eq. 18) rewards high cosine similarity between connected node embeddings, and NLSC (Eq. 19) penalizes high similarity of non-connected nodes. These operations are beneficial only when connected nodes tend to share attributes or labels. On heterophilous graphs—common in SIoT settings where edges often connect complementary device types or different social roles—the same operations spread incorrect attributes across edges and actively distort the embedding space. The experimental section evaluates only Cora, Citeseer, PubMed, Amac, Amap, and CS, all standard homophilous citation or co-purchase graphs; no SIoT or heterophilous dataset appears. The paper's conclusion that TDAR provides 'a robust solution to the challenges posed by AMGs' in SIoT is therefore unsupported for exactly the motivating application domain. This is a correctness/scope risk, not merely a preference for a different benchmark family.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TDAR, a framework for learning from graphs with missing node attributes (AMGs). TDAR combines a topology-aware attribute propagation pre-filling step (TAAP) with embedding-space confidence weighting (ESPC) and two regularization terms (NHS and NLSC) inside a graph autoencoder, and is motivated by attribute-missing graphs in the Social Internet of Things. Experiments on six standard citation and co-purchase datasets report improvements over existing AMG baselines in attribute reconstruction, node classification, and clustering, along with ablations, hyperparameter studies, and qualitative visualizations. The authors provide a link to the code.","tokens_in":24767,"tokens_out":11436,"duration_ms":106484,"significance":"If validated, TDAR would offer a useful end-to-end method for attribute-missing graph learning, and the idea of combining pre-filling with topology-aware weighting and homophily regularizers is reasonable. The paper's strengths are its broad comparison with many baselines, the ablation of each component, the hyperparameter sensitivity analysis, and the release of code. However, the current evidence is weakened by the absence of any statistical variability analysis (all runs use a single fixed seed), by internal inconsistencies in the equations that define the core mechanisms, and by an experimental scope that excludes the heterophilous or SIoT-style graphs that the paper itself identifies as the motivating application domain. The improvements are consistent in direction but the claimed 'significance' is not established.","major_comments":[{"comment":"All reported results come from a single fixed random seed (72), with no error bars, multiple runs, or significance tests. Since many reported gains over the strongest baselines are small (1–4% on reconstruction and classification), the claim in the abstract and Section VI that TDAR 'significantly outperforms' state-of-the-art methods is not statistically supported. Please report results over multiple seeds with mean and standard deviation, perform significance tests where appropriate, or explicitly temper the strength of the claims.","section":"§V.C, Tables II–V"},{"comment":"The three core mechanisms—TAAP (Eq. 8), NHS (Eq. 18), and NLSC (Eq. 19)—explicitly encode a homophily and smoothness prior: features are propagated along edges, connected nodes are pushed to have similar embeddings, and non-connected nodes with similar embeddings are penalized. The experimental evaluation, however, is restricted to homophilous citation graphs (Cora, Citeseer, PubMed) and co-purchase graphs (Amac, Amap, CS), with no heterophilous dataset and no SIoT dataset. Section V.G measures homogeneity as a desirable outcome rather than testing robustness to heterophily. Given the paper's SIoT motivation and its claim of 'a robust solution' (abstract, Section VI), the evidence does not support the robustness claim for the motivating domain. Please add experiments on heterophilous or SIoT-style graphs, or restrict the conclusion to homophilous AMGs.","section":"§IV.C and §V.A/V.G"},{"comment":"The confidence weight matrix W in Eq. (11) is not well-defined. The formula 'W = α F_{k2u}^T + (1 − α F_{u2k}^T)' mixes matrices of incompatible dimensions (F_{k2u} and F_{u2k} are introduced as N×D expansions, so their transposes are D×N), and the second term is ambiguous. Moreover, the described monotonicity is contradicted: the text says unknown-node weights should decrease with distance to known nodes, but the formula uses a positive multiple of distance, and the known-node term '1 − α F_{u2k}^T' would decrease (and possibly become negative) as the number of unknown neighbors grows, contrary to the statement that it should increase. Because the ESPC module is a central contribution, this formula must be corrected and clarified before the method can be reproduced or fairly evaluated.","section":"§IV.B, Eq. (11)"},{"comment":"The TAAP update is internally inconsistent. Eq. (7) writes the update as a propagation matrix plus a vector [α X_k^{(l)}; (1−β)X_k^{(0)}], but Eq. (8) updates unknown nodes with an added term α X_k^{(l)} (a mean feature vector) and updates known nodes through the two-step form X_k^{(l+1)} = \\hat{A}_{ku}X_u^{(l)} + \\hat{A}_{kk}X_k^{(l)}, \\tilde{X}_k^{(l+1)} = X_k^{(0)} + β X_k^{(l+1)}. These expressions do not match algebraically, and the second block of the added vector in Eq. (7) appears to have length k rather than the required unknown-node count. The actual algorithm implemented is therefore ambiguous, which is a serious reproducibility problem for the pre-filling step that the paper identifies as a key contribution.","section":"§IV.A, Eqs. (7)–(8)"},{"comment":"ITR and MATE are reported as out-of-memory on PubMed and CS, which removes two of the strongest baselines from the node classification comparison on the two largest datasets. The very large claimed improvements on those datasets (e.g., 32.1% on PubMed and 9.5% on CS) are therefore not established against the full set of competing state-of-the-art methods. Please provide results for memory-efficient runs of these baselines, or clearly discuss the missing comparisons and their effect on the 'consistently outperforms' conclusion.","section":"Table IV"}],"minor_comments":[{"comment":"The baseline row labeled 'GGN' appears to be a typo for GCN; please correct it. Also, 'GNN*' is used as a single aggregated row in Tables II, IV, and V, but GCN, GraphSAGE, and GAT are listed as separate rows in Table III; please unify the presentation.","section":"Table III"},{"comment":"'Amcp' is a typo for 'Amap', and the phrase 'binary-tier GCN-MLP' is unclear; please specify the exact encoder/decoder architecture.","section":"§V.C"},{"comment":"The metric name 'nDGC@k' in the description of Figure 4 should be 'nDCG@k'.","section":"Figure 4 text"},{"comment":"The symbol α is used for two different hyperparameters: the TAAP global propagation coefficient (Section IV.A, set to 0.05) and the ESPC distance attenuation factor (Section IV.B, later set to 0.9 in §V.C). Please use distinct symbols to avoid ambiguity and to make the hyperparameter reporting unambiguous.","section":"§IV.A and §IV.B"},{"comment":"Reference [49] cites Shannon's 'A Mathematical Theory of Communication' as the basis for Dirichlet energy minimization in graphs; this appears to be a mis-citation. Please replace it with an appropriate graph Laplacian or harmonic function reference, e.g., Zhu et al. (ICML 2003).","section":"Reference [49]"},{"comment":"The sentence 'TDAR method consistently outperforms the other methods across all four datasets' does not match Table IV, which contains six datasets; please correct the count.","section":"§V.D.2"},{"comment":"The complexity summary says the total complexity 'simplifies to O(N^2)', but the NLSC term is already O(N^2) (more precisely O(N^2 D) for pairwise cosine similarities over non-edges), so this is not a simplification but the dominant term. Please state the dependence on D explicitly and clarify the O(N log N) estimate for NHS, which neglects the per-edge inner product cost.","section":"§IV.E"},{"comment":"In the correlation matrix C, the indices i and j are used for feature dimensions while v is used for nodes, but \\bar{Z}_i and \\bar{Z}_j are not explicitly defined as means over nodes; please clarify the notation.","section":"Eq. (12)"},{"comment":"The reconstruction loss L_TAAP supervises against \\tilde{X}_k, the TAAP-refined known attributes, rather than the original X_k. Since TAAP modifies known attributes with a β-weighted propagation term, the loss does not directly enforce fidelity to the original known values; please justify this design choice.","section":"Eq. (17)"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to target an SIoT/IoT journal, but the experiments contain no SIoT-specific data and the claims of robustness for SIoT are not supported by the evaluation. The internal inconsistencies in Eqs. (7), (8), and (11) are the most serious technical issues and need to be fixed before the method can be assessed. The lack of any multi-seed statistics is also a substantive gap for a journal submission. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: TDAR is a competent, incremental contribution to attribute-missing graph learning. The gains are real on the benchmarks tested, but the paper's SIoT framing and \"robust solution\" claim outrun the evidence. The method is an integration of feature propagation, confidence weighting, and homophily regularizers, each familiar; the integration and the strong benchmark comparison are the value.\n\nWhat's actually new: TAAP modifies Feature Propagation with a global mean term and a known-reset, ESPC uses distance-to-known and known-node-unknown-neighbor counts to reweight embeddings, and NHS/NLSC add homophily and negative-similarity losses. None of these is a conceptual leap, but the full pipeline is new and the ablations show each piece contributes. Code is available. On six standard datasets, TDAR consistently outperforms SAT, SVGA, ITR, MATE, AIAE on reconstruction, classification, and clustering, with gains around 1-5% on reconstruction and larger on clustering. That is a useful result for practitioners.\n\nSoft spots: (1) All results are from one seed (72), no error bars, no significance tests. The gains are small enough that this matters. (2) The homophily assumption is baked into every component: TAAP propagates known features along edges, NHS pulls connected embeddings together, NLSC pushes unconnected apart. The paper only tests homophilous citation and co-purchase graphs. For SIoT graphs, where edges may connect complementary device types, these same mechanisms could spread wrong attributes. The conclusion that TDAR is 'a robust solution' for AMGs in SIoT is not supported without at least one heterophilous or real SIoT dataset. (3) Minor: Table III has 'GGN' for GCN, ITR and MATE OOM on PubMed/CS, and the TAAP equations (7)-(8) have a notational mismatch between the Theta vector and the actual updates; the Dirichlet-energy motivation is a heuristic, not a derivation.\n\nBottom line: The central empirical claim stands for the tested distribution. The paper is a solid within-subfield advance, not a breakthrough. It deserves a serious referee, but the referee should ask for multi-seed results and either a heterophilous experiment or an honest limitation paragraph.","headline":"Solid incremental integration of known ideas with consistent benchmark gains, but the SIoT robustness claim needs a heterophilous test.","tokens_in":25287,"tokens_out":3336,"would_cite":true,"duration_ms":32617,"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":"TDAR, a topology-driven framework, recovers missing node attributes in graphs by pre-filling with edge-based propagation and regularizing embeddings, and reports state-of-the-art results on reconstruction, classification, and clustering…","keywords":["attribute missing graph","feature propagation","graph autoencoder","attribute recovery","homophily","node classification","graph clustering","social internet of things"],"falsifier":"On a heterophilous graph (edges mostly between dissimilar nodes) with 50% masked attributes, compare TDAR's reconstruction RMSE with a plain MLP autoencoder that ignores graph structure; if the MLP is more accurate, topology-driven pre-fill and homophily regularizers are actively harmful in that regime.","tokens_in":24354,"feed_emoji":"🔗","tokens_out":7608,"duration_ms":67228,"temperature":0.7,"pith_summary":"The paper claims that in graphs where some nodes have missing attributes, the graph's own topology is enough to drive high-quality attribute recovery, without needing generative adversarial training or complex probabilistic models. It introduces TDAR, a framework built on a graph autoencoder that (1) pre-fills missing attributes by propagating known ones along edges under a Dirichlet-energy-minimization view, (2) re-weights the learned embeddings according to each node's topological position (distance to known nodes for unknown nodes, number of unknown neighbors for known nodes), and (3) adds two regularizers that pull connected nodes together in embedding space while suppressing misleading similarities between disconnected nodes. Across six citation and product graphs, TDAR reports consistent state-of-the-art results on attribute reconstruction, node classification, and clustering, with average reconstruction gains around 3% and substantially larger gains on clustering.","feed_headline":"Graph topology recovers missing node attributes, beating prior methods","feed_subtitle":"Pre-filling via edge propagation and embedding regularizers lifts clustering accuracy by up to 35%.","key_machinery":"The load-bearing identity is the Dirichlet energy $E(X)=X^T L X$ of the graph Laplacian, whose minimizer yields the feature-propagation update $X_u = -L_{uu}^{-1} L_{uk} X_k$. TDAR's key mechanism is a modification of the iterative feature-propagation loop (Eqs. 7–8) that injects a small global average of known attributes and lets known attributes participate in updates with a reset term, cutting the required iterations from roughly 40 to at most 10 and reducing oversmoothing. On top of that, ESPC computes a per-node confidence weight from two distance functions—shortest path from unknown nodes to known nodes, and known nodes' count of unknown neighbors—and applies it to a correlation-corrected embedding to re-weight the loss; NHS and NLSC then add cosine-similarity losses over connected and disconnected node pairs.","core_discovery":"On its own terms, TDAR establishes that the central difficulty of learning from attribute-missing graphs is not encoding capacity but initialization and supervision: how the missing attributes are filled before network training, and how the embedding space is constrained. The paper's discovery is that a cheap, parameter-free pre-fill step—feature propagation with a small global-mean term and a known-reset mechanism—already outperforms the standard FP pre-fill, and that combining it with position-aware embedding weights (ESPC) and two similarity-based regularizers (NHS and NLSC) yields the best reported results on all benchmark datasets. The authors argue that this shows topology should be treated as the primary signal for attribute recovery, with generative refinements playing a secondary role.","pith_inferences":["If the homophily assumption is the real source of gains, TDAR should degrade on heterophilous graphs (e.g., protein interaction or transaction networks); a direct test on such data would mark the boundary of the method.","The NHS loss pushes every connected pair toward similarity, which may over-constrain hubs and noisy-edge graphs; a degree-normalized or attention-weighted variant of NHS could preserve the benefit while avoiding collapse.","The global-mean term in TAAP injects a dataset-wide average into every unknown node; on highly imbalanced label distributions this could bias recovery toward majority classes, and ablating α's value at larger sizes would reveal whether this matters.","The reported clustering gains on Amap are much larger than on Cora; investigating whether this comes from the graph's dense structure or from the pre-fill would tell practitioners which datasets most benefit from TDAR."],"forward_implications":["Graphs with missing attributes can be handled with a deterministic, parameter-free pre-fill plus regularized GAE training, avoiding GAN and variational inference overhead.","The improvements hold across missing rates from 0.2 to 0.8, so topology-driven recovery remains useful even under extreme missingness.","Better reconstruction transfers to downstream tasks: node classification accuracy improves by up to 9.5 points and clustering accuracy by up to 35 points on the tested benchmarks.","The overall complexity stays $O(N^2)$, dominated by the non-link similarity calibration, keeping TDAR competitive in runtime with prior state-of-the-art methods.","The framework's modular design means each component (pre-fill, weighting, regularizers) can be lifted into other AMG pipelines as an initialization or loss term."],"supporting_citations":[{"why":"Supplies the Feature Propagation method based on Dirichlet energy minimization that TAAP modifies and improves.","marker":"[15]"},{"why":"Extends FP with pseudo-confidence weighting; ESPC's distance-based confidence weights develop this idea.","marker":"[50]"},{"why":"ITR, the initialize-then-refine baseline that TDAR builds upon and outperforms; motivates the pre-fill strategy.","marker":"[33]"},{"why":"SVGA, the structured variational graph autoencoder baseline that TDAR compares against on reconstruction and downstream tasks.","marker":"[29]"},{"why":"MATE, a multi-view graph imputation baseline that is among the strongest recent competitors in TDAR's experiments.","marker":"[34]"},{"why":"AIAE, an attribute imputation autoencoder baseline that represents the recent state of the art in AMG learning.","marker":"[35]"},{"why":"VGAE/GAE base autoencoder architecture that TDAR's encoder-decoder and reconstruction loss derive from.","marker":"[25]"}],"fun_headline_variants":["Topology recovers missing node attributes, beating prior methods","TDAR pre-fills attributes via topology, boosts clustering accuracy","Attribute-missing graphs: topology pre-fill outperforms SOTA","Topology-driven attribute recovery lifts downstream task performance","Pre-fill with graph topology recovers attributes, beats baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes homophily—that connected nodes have similar attributes—so feature propagation along edges and the NHS regularizer improve recovery; on graphs where edges connect dissimilar nodes or are noisy, these steps spread incorrect attributes and can hurt reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["Topology recovers missing node attributes, beating prior methods","TDAR pre-fills attributes via topology, boosts clustering accuracy","Attribute-missing graphs: topology pre-fill outperforms SOTA","Topology-driven attribute recovery lifts downstream task performance","Pre-fill with graph topology recovers attributes, beats baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000334,"raw_usage":{"total_tokens":1825,"prompt_tokens":887,"completion_tokens":938,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":854}},"tokens_in":503,"tokens_out":938,"duration_ms":9360,"temperature":1.0,"reasoning_tokens":854,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:23:29.624067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a heterophilous graph (edges mostly between dissimilar nodes) with 50% masked attributes, compare TDAR's reconstruction RMSE with a plain MLP autoencoder that ignores graph structure; if the MLP is more accurate, topology-driven pre-fill and homophily regularizers are actively harmful in that regime.","supporting_citations":[{"cited_title":"On the unreasonable effectiveness of feature propagation in learning on graphs with missing node features,","cited_arxiv_id":null,"evidence_quote":"Supplies the Feature Propagation method based on Dirichlet energy minimization that TAAP modifies and improves."},{"cited_title":"Confidence-Based Feature Imputation for Graphs with Partially Known Features","cited_arxiv_id":"2305.16618","evidence_quote":"Extends FP with pseudo-confidence weighting; ESPC's distance-based confidence weights develop this idea."},{"cited_title":"Initializing then refining: A simple graph attribute imputation network,","cited_arxiv_id":null,"evidence_quote":"ITR, the initialize-then-refine baseline that TDAR builds upon and outperforms; motivates the pre-fill strategy."},{"cited_title":"Accurate node feature estimation with structured variational graph autoencoder,","cited_arxiv_id":null,"evidence_quote":"SVGA, the structured variational graph autoencoder baseline that TDAR compares against on reconstruction and downstream tasks."},{"cited_title":"Multi-view graph imputation network,","cited_arxiv_id":null,"evidence_quote":"MATE, a multi-view graph imputation baseline that is among the strongest recent competitors in TDAR's experiments."},{"cited_title":"Attribute imputation autoencoders for attribute-missing graphs,","cited_arxiv_id":null,"evidence_quote":"AIAE, an attribute imputation autoencoder baseline that represents the recent state of the art in AMG learning."}],"review_version":1}