acceptodds
Under review as a conference paper at ICLR 2027

Selective Feature-Node Transformation for Heterophilic Graph Learning

Abstract

Graph neural networks (GNNs) can struggle on heterophilic graphs, where neighborhood aggregation mixes information from nodes with different labels. Feature-node transformation addresses this challenge by connecting original nodes through shared discrete features. However, exhaustive transformation includes every active feature, regardless of its predictive relevance or structural cost. This work proposes a selective transformation framework that identifies useful auxiliary structure under a feature-node budget while retaining the original attributes and topology. A relevance score combines normalized mutual information, pairwise label agreement above chance, coverage, and incidence-edge cost. Support-based shrinkage and component standardization balance these signals when labeled observations are limited. The selected incidence graph is integrated with the original topology through signed, feature-aware gated propagation. Analysis establishes exact auxiliary graph sizes and a sufficient condition for above-chance agreement along selected two-hop paths. Evaluation on six node-classification benchmarks shows substantial reductions in auxiliary nodes and edges, accompanied by lower runtime and memory measurements. Predictive outcomes vary across datasets, while matched-budget comparisons and ablations clarify when relevance scoring helps and when simpler criteria remain competitive. These findings establish selective feature-node construction as a practical approach to balancing predictive performance and computational cost in graph learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
heterophilic graph learning, graph neural networks, graph transformation, feature-node selection, semi-supervised node classification, efficient graph learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_bBrwOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a feature-node transformation method to selectively convert discrete features into auxiliary message-passing nodes for heterophilic graph learning, building directly on the exhaustive feature-node transformation approach GRAHITE (Qiu,2026). The proposed method ranks candidate features under a node budget K using a composite heuristic score. The authors provide bounds on graph size and two-hop label agreement, and evaluate the approach across six node-classification benchmarks (4 heterophilic,2 homophilic).

Critical strengths
  • The motivation behind is practical. To tackle computational overhead of exhaustive feature-node construction as in GRAPHITE is reasonable.
  • The experimental results demonstrate the computational efficiency ot the proposed method. As shown in Tab.2&5, restricting the number of auxiliary feature nodes predictably reduces the number of added incidence edges, leading to lower runtime and memory consumption.
Critical weaknesses
  • Limited technical novelty. The methodological contribution is highly incremental. First, the core idea of transforming discrete features into auxiliary nodes is directly inherited from GRAPHITE, while the selection mechanism is essentially a hand-crafted linear combination of standard statistical metrics (ie., mutual info., class pair purity, log-freq., and degree cost), coupled with z-score standardization and basic empirical shrinkage. Second, the theoretical results in Proposition 1 and Theorem 1 are mathematically trivial. Proposition1 merely counts added nodes and edges, while Theorm 1 simply states that a convex combination of terms greater than p0p_0 is also greater than p0p_0.

  • Empirical results are unconvincing and contradictory with the paper's premise. First of all, although the paper is explicitly framed around heterophilic graphs, Table 2 shows that selective construction degrades acc. across all four heterophilic datasets compared to the controlled baseline, and only marginally improve on two homophilic datasets. Moreover, the results in table 3 show the inferiority of the proposed method to GRAPHITE and standard baselines (underperforming GRPAPHITE on 5/6 benchmarks, including a massive drop from 94.78% to 86.22% ROC-AUC on Minesweeper). A more concerning drop is on the homophilic graphs, the proposed method performs noticably worse than a vanilla GCN as well as GAT and GraphSAGE.

  • The composite score looks ineffective, based on Table 4. When comparing the proposed relevance score against naive baselines under the same node budget, the proposed method fails to outperform simple heristics on 3 out of 4 heterophilic graphs. In particular, 'Random' and Mutual Info. selection outperform the proposed method on Squirrel, while simple Frequency selection outperforms selective method on two datasets. This calls into question the necessity of the multi-component scoring (Eq.13).

  • Quite superficial experimental analysis. The discussion of experiment resutls is overly brief and lacks depth. No substantive analysis of why the proposed method outperform or lags behind strong baselines, no any core claims and takeaways from Table 3. Besides, framing Table 3 as "Architecture context" does not excuse the lack of rigorous comparison and explanation for the performance drops.

Severe Reproducibility issues: There are some weird (sounds alarming) statements iregarding the experimental integrity and reproducibility n the paper, e.g., in the caption of Table 5 (Line 397), the authors state that "measurement scope and diagnostic label masks require verification", and Line 466-470 (Reproducibilit paragraph): " Exact relication remain limited by unavailable code, seeds, ..., ablation error-bar defintion are also unspecified." Furthermore, the hyperparmeters like λr,g,v,e\lambda_{r, g, v,e} in Eq.13 are never specified in the text.

Questions for authors

Please find the concerns in the ''weakness"

2 comments
AuthorsOpenReview ↗

Official Comment to Reviewer bBrw

Response to Reviewer bBrw

We thank the reviewer for the critical assessment and for acknowledging the practical motivation and measured computational reductions. We agree that the methodological novelty, empirical trade-offs, and reproducibility require clearer justification.

1. Novelty relative to GRAPHITE.

We acknowledge that feature-node transformation itself originates from GRAPHITE and that the individual statistics in Eq. (13) are established concepts. The proposed novelty is not a new mutual-information estimator, graph transformation primitive, or propagation architecture.

The distinction is the structural selection problem: given a fixed auxiliary-node budget, which features should become additional message-passing nodes, considering both predictive relevance and the number of incidences introduced?

This differs from conventional input-feature selection because all original attributes and topology remain intact. Selection controls additional connectivity rather than removing input information. Equations (2)–(3) formulate this decision, while Eqs. (9)–(13) provide a practical support-adjusted surrogate combining class association, feature-sharing agreement, coverage, and incidence cost.

We agree that this is an extension of feature-node transformation and will emphasize its specific contribution without claiming novelty for the underlying statistical components.

2. Theoretical contribution.

The reviewer correctly notes that Proposition 1 follows from incidence counting and Theorem 1 from a weighted-average identity. These results are not intended as deep learning-theoretic guarantees. They establish exact construction costs and characterize labeled two-hop path agreement after selecting a feature subset.

Their purpose is to separate what can be guaranteed structurally from what must be established empirically. In particular, neither the ranking rule nor higher path agreement guarantees improved classification. We will ensure these claims are presented at their appropriate theoretical scope.

3. Predictive performance and GRAPHITE comparison.

We agree that selective construction does not consistently improve predictive accuracy. However, Table 3 contains published architecture results rather than controlled reruns, so the published GRAPHITE values cannot isolate the effect of feature selection.

This distinction is especially important for Minesweeper. Published GRAPHITE achieves 94.78 ROC-AUC, whereas the controlled exhaustive implementation achieves 87.33 and selective construction achieves 86.22. Thus, the controlled selection-associated difference is 1.11 points, not the 8.56-point difference between the published GRAPHITE result and the selective implementation. The discrepancy between the published and controlled exhaustive results requires investigation and cannot be attributed to selection.

Similarly, Table 3 does not establish state-of-the-art predictive performance. The intended contribution is an explicit structural-budget mechanism and its measured efficiency–accuracy behavior, rather than superiority over every GNN architecture.

4. Necessity of the composite score.

We agree that Table 4 does not demonstrate universal superiority over simple heuristics. The combined score achieves the highest mean on ACTOR, but frequency or other selectors can be preferable elsewhere. Figure 2 similarly shows that individual components are not uniformly beneficial.

These findings suggest that structural relevance depends on the graph and operating budget. The combined score is a configurable mechanism rather than an empirically established universally optimal selector. Equal-incidence-cost comparisons and paired evaluations are necessary before stronger conclusions can be drawn.

5. Reproducibility and experimental integrity.

We agree that the missing implementation details are serious limitations. In particular, the score weights in Eq. (13), optimizer settings, seeds, hardware configuration, measurement definitions, and ACTOR feature-count discrepancy require verification.

The statements in Section 4.5 were included to disclose these limitations, not to imply that the experimental record had been fully verified. Nevertheless, acknowledging missing information does not substitute for providing it.

We will prioritize recovery and documentation of the original experimental configurations, correct any verified inconsistencies, and ensure that unsupported claims are not retained. The exact graph-size identities are independent of these implementation details, whereas the measured runtime, memory, and predictive results require their proper documentation.

We appreciate the reviewer's assessment of the practical problem. We hope these clarifications help distinguish the proposed structural selection contribution from inherited feature-node transformation and establish a more precise interpretation of the empirical evidence.

I thank the authors for their detailed response, which addresses some of my concerns regarding the experiments. I will therefore raise my score to 2. However, I remain concerned that the paper’s technical contribution is limited.

Reviewer_cySMOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a question arising from feature-node transformations for heterophilic graph learning: if converting discrete features into auxiliary nodes is useful but expensive, which features should be selected?

The proposed method retains the original graph and all node attributes, while constructing auxiliary nodes only for a selected subset of features. Selection uses a score combining mutual information with class labels, above-chance label agreement among feature-sharing nodes, coverage, incidence cost, and a support-based reliability adjustment. A signed gated propagation model then operates on the resulting augmented graph.

The analysis characterizes graph size and conditional agreement along feature-induced paths. Experiments on six node-classification benchmarks compare exhaustive and selective construction, simpler selection strategies, and computational costs.

Critical strengths

S1: The paper studies a question arising from feature-node transformations for heterophilic graph learning: if converting discrete features into auxiliary nodes is useful but expensive, which features should be selected? The method retains the original graph and all node attributes, while constructing auxiliary nodes only for a selected subset of features. Selection uses a score combining mutual information with class labels, above-chance label agreement among feature-sharing nodes, coverage, incidence cost, and a support-based reliability adjustment. A signed gated propagation model then operates on the resulting augmented graph.

S2: This is an appreciated aspect of the paper. The theoretical analysis does not claim that selecting feature nodes guarantees better classification, and the experiments report cases in which the retained paths exhibit stronger label agreement while accuracy declines. The controlled exhaustive-versus-selective comparison is also more informative than presenting only a collection of published GNN baselines.

Critical weaknesses

W1: This is my main concern. In Table 2, selective construction decreases the mean predictive score on all four heterophilic benchmarks. On Chameleon-F, for example, accuracy drops from 48.93 to 44.72 while 88.9% of the auxiliary edges are removed. Such a trade-off might be acceptable if the paper demonstrated that its score produces a superior accuracy-efficiency frontier.

W2: Proposition 1 follows directly from counting selected feature nodes and incidences. Theorem 1 is mathematically reasonable, but its sufficient condition for above-chance two-hop agreement is essentially a weighted-average property of the selected features. Crucially, the proposed top-K rule does not guarantee the condition, and the result provides no direct information about predictive generalization. I do not regard this as a correctness problem. Rather, the theoretical results establish structural properties that are weaker than the central algorithmic question of whether the proposed selection improves learning under a budget. The evaluation would also benefit from greater reproducibility. Section 4.5 acknowledges missing implementation details and unspecified aspects of the memory, runtime, and diagnostic measurements. These uncertainties make it harder to independently assess the reported efficiency gains.

W3: As a side note, I think the authors cherry-picked the benchmarks and mostly focused on old and tortured (research-wise) benchmarks where research has been squeezing decimals in the last years. Frankly, these benchmarks are ridiculously too small! The authors did not provide any sensible real-world usage of the technique.

Questions for authors

Q1: At equal incidence-edge budgets, rather than equal numbers of feature nodes, how does the proposed score compare with frequency, mutual information, and random selection?

Q2: Can the authors quantify whether the combined relevance score provides consistent improvements across matched splits, rather than comparing only aggregate means?

Q3: Could the authors clarify the label mask used for the path-agreement diagnostics and the precise scope of runtime and memory measurements? This matters for interpreting the relationship between the theoretical analysis and the experimental results.

Q4: There are much larger datasets to consider. I wonder why the authors focused on such small datasets

2 comments
AuthorsOpenReview ↗

Official Comment to Reviewer cySM

Response to Reviewer cySM

We thank the reviewer for the detailed assessment and for recognizing the motivation behind selective feature-node construction and the distinction between structural analysis and predictive generalization. We agree that the accuracy–efficiency trade-off needs stronger empirical characterization.

1. Predictive performance and the accuracy–efficiency trade-off (W1, Q1–Q2).

The reviewer correctly observes that selective construction reduces mean predictive performance on all four heterophilic benchmarks relative to exhaustive construction. This is an important result rather than an outcome we intend to obscure.

For example, on Chameleon-F, selective construction removes 88.9% of auxiliary incidence edges while decreasing accuracy from 48.93% to 44.72%. Conversely, on CiteSeer, it removes 91.8% of auxiliary edges while increasing accuracy from 72.79% to 74.19%. These findings demonstrate that the value of additional auxiliary connectivity depends on the dataset and operating budget.

The combined score considers both label relevance and structural cost, but we acknowledge that Table 4 establishes neither a superior accuracy–cost frontier nor consistent statistical improvements over simpler selectors. In particular, the current comparisons use equal auxiliary-node budgets, not equal incidence-edge budgets.

We agree that matched-edge-budget comparisons and split-level paired analysis are important to establish the additional value of joint scoring. These conclusions cannot be inferred from the published aggregate statistics alone, and we will avoid making unsupported superiority or significance claims.

2. Significance of the theoretical analysis (W2).

We agree that Proposition 1 and Theorem 1 are deliberately modest results. Their role is to characterize the consequences of selection, not establish a new generalization bound.

Proposition 1 gives exact structural costs for any selected feature subset. Theorem 1 characterizes when the retained feature-induced paths have above-chance label agreement. Importantly, Eq. (13) does not guarantee the theorem's sufficient condition.

The distinction is empirically relevant: on Chameleon-F, selected-path agreement increases from 0.2588 to 0.5968, while accuracy declines. This demonstrates why path agreement cannot be substituted for predictive utility. We will further clarify this limited theoretical scope.

3. Benchmark scale and practical relevance (W3, Q4).

We acknowledge the limitation regarding dataset scale. The six benchmarks were selected to evaluate the proposed transformation under established node-classification protocols and to enable comparison with related heterophily research, including GRAPHITE. However, they contain at most 10,000 original nodes, and the current results do not establish large-scale effectiveness.

The structural analysis applies to larger graphs, with construction cost dependent on the number of retained incidences, but empirical scalability remains an open question. We agree that larger real-world graphs would strengthen the practical evidence and will make this limitation explicit.

4. Reproducibility and diagnostic measurements (Q3).

Equations (1), (11), and (15) define label statistics using the training-label set LL. However, the actual label mask used to generate Table 5's diagnostics must be confirmed from the experimental records. Similarly, the timing interval, memory definition, and hardware details require verification. We will not interpret these measurements as evidence of held-out generalization or hardware-independent speedup.

We appreciate the reviewer's recognition that the controlled exhaustive comparison provides useful evidence. We hope the distinction between budgeted structural efficiency, conditional path properties, and predictive performance clarifies the contribution and its limitations.

Reviewer_cySMOpenReview ↗

Narrow contribution

Thank you for taking the time to read my comments and provide some potential solutions. Nevertheless, the main limitations of the paper- dataset size and narrow theoretical results are not adequately addressed in the rebuttal (I did not expect that this could have happened).

Overall, the paper does not provide enough evidence to justify an acceptance.

Reviewer_ZNBjOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes a selective feature-node transformation for heterophilic graph learning. It ranks features using label relevance, pairwise agreement, coverage, and incidence cost, then constructs auxiliary nodes under a fixed budget while preserving original attributes and edges.

Critical strengths

This work provides explicit control over graph expansion without discarding original attributes or topology, addressing a practical limitation of exhaustive transformation.

Critical weaknesses

1- Key methodological concepts are insufficiently defined or justified. These gaps obscure how the method operates and prevent a reliable assessment of its validity.

2- The “Predictive and Structural Evaluation” section lacks essential implementation and experimental details. This limits reproducibility and makes it difficult to determine whether the reported results substantiate the method’s effectiveness.

Questions for authors
  • Please define dd when it first appears in line 127.
  • How is the auxiliary node’s positive status determined and justified? Please visualize the dynamic graph construction process.
  • The notation V∗SV*S in Figure 1 is unclear. Please use standard terminology and define it consistently.
  • Please explain the purpose of the tanh⁡\tanh function and justify its use.
  • The “Predictive and Structural Evaluation” section lacks sufficient implementation and experimental details for reproducibility.
3 comments
AuthorsOpenReview ↗

Official Comment to Reviewer ZNBj

Response to Reviewer ZNBj

We thank the reviewer for identifying points where the presentation needs greater precision. We agree that clear definitions and implementation details are essential for evaluating the proposed method. We address each question below.

1. Definition of dd and graph notation.

In Section 3.1, dd denotes the number of original input-feature dimensions in X∈{0,1}n×dX\in\{0,1\}^{n\times d}. In contrast, D=∣F∣D=|\mathcal F| denotes the number of active feature candidates, and K≤DK\leq D is the number selected for auxiliary construction. We will state this distinction explicitly.

The notation VS∗V_S^* represents the augmented vertex set, defined in Eq. (5) as VS∗=V∪{xk:k∈S}V_S^*=V\cup\{x_k:k\in S\}. Thus, the transformed graph contains all original nodes plus the selected auxiliary feature nodes. We agree that Figure 1 should use and explain this notation consistently.

2. Auxiliary-node activation and construction.

An auxiliary node is not assigned a positive class label. Instead, a binary feature is active at original node ii when Xik=1X_{ik}=1. As described in Section 4.1, nonnegative attributes are binarized using positive values, while attributes containing negative values use the feature-column mean as the threshold.

The construction follows five steps: compute feature statistics using available attributes and training labels; rank active features; select the top KK; create one auxiliary node xkx_k per selected feature; and connect xkx_k to every original node satisfying Xik=1X_{ik}=1. Equations (5)–(6) define the resulting adjacency and auxiliary features. The original graph and all attributes remain unchanged. Selection is performed before model training; it is not a continuously changing temporal graph. We will make this sequence more explicit in Figure 1 and its explanation.

3. Purpose of tanh⁡\tanh.

The tanh⁡\tanh function in Eq. (7) maps learned gate values to [−1,1][-1,1], allowing a message to contribute positively or negatively, unlike a nonnegative sigmoid gate. The receiver-first formulation also permits asymmetric message weighting. The temperature τ\tau controls gate saturation. This propagation mechanism integrates the constructed graph and is not claimed as an independent novelty.

4. Experimental reproducibility.

We acknowledge that Section 4.1 and Table 1 do not fully specify the implementation. In particular, the score weights, training configuration, software and hardware information, and exact measurement conventions require documentation. We will distinguish parameters already specified in the manuscript from settings that still require verification against the experimental records.

The principal contribution is selective auxiliary graph construction under an explicit budget, rather than the introduction of an entirely new feature-node transformation. We hope these explanations resolve the conceptual ambiguities and make the methodological contribution easier to assess.

Reviewer_ZNBjOpenReview ↗

Keeping my recommendation marginal

The rebuttal regarding the implementation and experimental details for reproducibility remains insufficient. Moreover, the authors should explain the motivation for using a specific function or strategy, rather than just defining the mathematics.

AuthorsOpenReview ↗

Methodological Design Choices and Experimental Reproducibility

We thank the reviewer for the continued feedback. We agree that methodological choices require clear technical justification and that reproducibility depends on sufficiently detailed experimental settings.

The proposed scoring components address complementary aspects of feature-node construction. Mutual information captures feature–label dependence, while above-chance pairwise agreement estimates the relevance of feature-induced message-passing paths. Coverage favors broader connectivity, whereas incidence cost penalizes graph expansion. Support-based shrinkage mitigates unreliable estimates from limited labeled observations, and standardization balances the contributions of differently scaled statistics. Top-KK selection explicitly controls the auxiliary-node budget without removing original attributes or edges. Furthermore, the tanh⁡\tanh gate permits signed message contributions, allowing the propagation mechanism to accommodate potentially supportive or conflicting neighborhood information in heterophilic graphs.

Regarding reproducibility, Eqs. (5)–(13), the dataset protocols, and Table 1 establish the main methodological framework. We acknowledge that the exact scoring weights, optimization settings, random seeds, hardware configuration, and runtime and memory measurement procedures require further specification. We will clarify these details based on the verified experimental records and resolve the identified reporting inconsistencies in the revised manuscript. We appreciate the reviewer's constructive suggestions, which will help strengthen the methodological justification, reproducibility, and overall technical clarity of this work.

Reviewer_v45BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends exhaustive feature-node transformation by selecting which discrete features become auxiliary message-passing nodes while retaining the original attributes and edges. Its ranking combines mutual information, above-chance label agreement, coverage, and incidence cost, with support shrinkage and standardization. The analysis gives exact graph-size counts and a conditional result for agreement along feature-induced paths. Experiments on six datasets examine exhaustive versus selective construction, simple selectors, resource measurements, component ablations, and budget sensitivity.

Critical strengths
  1. The paper addresses a legitimate accuracy–cost tradeoff. Controlling auxiliary structure while retaining the original inputs is practically motivated. Table 2 reports 32.1–92.5% reductions in added incidence edges. Selective construction has lower means on the four heterophilic datasets and higher means on the two citation datasets; a predictive decrease at a substantially smaller structural budget can still be a useful operating point.

  2. The evaluation distinguishes several questions that should be assessed separately. The paper includes a controlled exhaustive comparison, random/frequency/MI selectors at matched node budgets, and budget curves. It distinguishes published baseline numbers from controlled reruns and states that top-K selection does not guarantee Theorem 1's premise. The conditional path-agreement result is limited, but the paper appropriately separates it from a guarantee of test accuracy.

Critical weaknesses
  1. The benefit of the combined ranking over simpler selection remains uncertain. In Table 4, the proposal has the highest mean on ACTOR and exceeds random selection's mean on three of four datasets. However, random and MI selection have means of 43.83 on Squirrel-F versus 43.59 for the proposal; frequency selection leads on Chameleon-F, 46.08 versus 44.89, and Minesweeper, 86.39 versus 86.09 ROC-AUC. These mean comparisons establish neither statistical differences nor equivalence. They leave the advantage over simple alternatives unclear, especially because equal node budgets need not have equal incidence costs. A comparison of accuracy against actual resource cost would better assess the ranking's benefit.

  2. The efficiency measurements need definitions to support a reproducible practical claim. Table 5 reports substantial reductions, including CiteSeer time from 127.73 to 22.78 seconds and memory from 15,830.8 to 2,572.0 MB. Section 4.5 leaves the timing interval, memory statistic, hardware metadata, diagnostic label mask, and ablation error-bar convention unspecified, along with several replication settings. It also identifies an ACTOR feature-count discrepancy. These omissions make the measurements difficult to interpret, particularly whether preprocessing and scoring are included. They do not demonstrate that the numbers are incorrect, but they weaken the efficiency evidence supporting the proposed ranking.

Questions for authors
  1. At matched incidence cost, or using measured total time and peak memory, does the proposed ranking offer a better accuracy–cost tradeoff than random, frequency, or MI selection? A focused comparison using the existing datasets and budget settings would address the main empirical uncertainty.
  2. Can you specify the timing scope, memory statistic, hardware and training settings, diagnostic label mask, and error-bar convention, and resolve the ACTOR count discrepancy? Please distinguish preprocessing and selection cost from model-training cost.
2 comments
AuthorsOpenReview ↗

Official Comment to Reviewer v45B

Response to Reviewer v45B

We thank the reviewer for the constructive assessment and for recognizing the practical motivation, controlled exhaustive comparison, and careful distinction between structural guarantees and predictive performance. We appreciate the two concerns and clarify them below.

1. Effectiveness of the combined relevance score.

The reviewer correctly observes that matching the number of auxiliary nodes does not ensure equal incidence-edge costs. Equation (13) is designed to balance predictive relevance and structural cost, rather than guarantee the highest accuracy at every budget. Table 4 shows that the combined score achieves the highest mean on ACTOR (37.36%), whereas simpler selectors remain competitive or outperform it on other datasets. We therefore do not claim universal superiority or statistical significance based on aggregate means.

The central contribution is the ability to make feature-node construction selective while preserving all original attributes and edges. Table 2 demonstrates 32.1–92.5% reductions in auxiliary incidences, and Table 5 reports lower measured runtime and memory across all six datasets. These results establish the practical motivation for budgeted construction, although they do not establish that the combined score achieves the best accuracy–cost frontier among selectors.

We agree that accuracy-versus-incidence-cost comparisons, using matched splits and consistent configurations, are the appropriate additional test. We will distinguish this outstanding comparison from the conclusions already supported by the existing experiments.

2. Efficiency measurements and reproducibility.

We agree that Table 5 requires more precise measurement definitions. The manuscript specifies training-label-only feature statistics, the construction procedure, dataset splits, and several validation-selected parameters, but does not sufficiently document timing boundaries, peak-memory definitions, hardware, optimizer settings, score weights, or diagnostic masks.

These omissions are limitations of the current reporting, not evidence of a theoretical efficiency guarantee. The exact structural reductions follow independently from Proposition 1; the measured time and memory differences remain implementation-dependent. The ACTOR feature-count discrepancy also requires verification against preprocessing records before correction.

We will prioritize verified implementation details and explicitly separate preprocessing, selection, graph construction, and model training wherever the original measurements permit.

We appreciate the reviewer's recognition that substantial structural savings at a modest predictive cost can constitute a meaningful operating point. We hope these clarifications help establish the intended contribution and its appropriately limited claims.

Reviewer_v45BOpenReview ↗

Response to rebuttal

Thank you for clarifying the scope. I agree that reducing auxiliary structure is valuable, but I still believe that clear measurement definitions and comparisons of accuracy versus incidence cost are important. After going over the authors’ responses and the other reviewers’ comments, I’ve decided to keep my rating unchanged.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.