acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Alignment-Driven Feature Augmentation for Graph Learning

Abstract

Graph Neural Networks (GNNs) have achieved strong performance on a wide range of graph learning tasks by integrating node attributes with graph structure through message passing. However, when structural and node attributes encode signals that are weak, inconsistent, or misaligned with the supervision objective, message passing can amplify noise and hinder performance. Existing approaches primarily address these challenges through architectural or structural modifications, and existing feature augmentation methods generally do not account for how the downstream GNN propagates the augmented signals. In this work, we adopt a data-centric perspective and propose GraphAlFA (Alignment-Driven Feature Augmentation), an adaptive, propagation-aware graph feature augmentation framework that explicitly enhances feature-structure-label alignment under task supervision. Rather than modifying model architecture or graph topology, GraphAlFA operates directly in the feature space by constructing complementary attribute-derived and structure-derived views, and optimizing them through complementary alignment objectives: supervised contrastive learning promotes label-discriminative representations, while a propagation-aware spectral objective estimates the downstream GNN's empirical frequency response and uses it to emphasize task-relevant spectral components effectively preserved by propagation. By conditioning augmentation on this empirical response instead of a fixed spectral prior, GraphAlFA adapts the input representation to the propagation behavior of the backbone while preserving the graph topology and backbone architecture. Theoretically, we provide an information-theoretic characterization of feature-label alignment and a propagation-aware characterization of spectral feature-structure alignment. Extensive experiments on synthetic and 9 real-world datasets spanning diverse homophily levels show that GraphAlFA consistently outperforms state-of-the-art baselines, achieving average accuracy gains of 4.82% on real datasets and 12.22% on synthetic datasets over original features.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Data Augmentation, Graph Neural Networks, Graph Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yYvdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to align feature, structure and label in graph learning. Specifically, the proposed method, GraphALFA, uses contrastive learning framework, where the two input views augments attribute and structure, followed by a learned gate for fusion. The model is optimized by two objectives, i.e., feature-lable and feature-structure alignments.

Critical strengths

None.

Critical weaknesses
  1. The authors combine several standard operations into one model. For example, the adopted attribute view, structure view and intra-/inter-class contrastive learning have been standard components in graph learning. Also, the used technique in structure view construction, (e.g., node2vec, LINE) are very old-fashioned.
  2. Potential high complexity. In feature-structure alignment loss computing, the authors firstly computed band-wise ratios of spectral energy change, then averaged them across layers for the final g_emp. That means we need to do spectrum-related operations every layer and every spectral band, which could incur high computing complexity.
Questions for authors
  1. In feature-structure alignment loss computing, it also requires label distribution over structure to be same as that of feature. So, could I view the two losses basically do the same thing that both align feature and label?
  2. This paper claimed the feature-structure-label alignment. But if the given graph natually does not satisfy this alignment, can we still learn such features to do the alignment?
  3. Lack the comparison with graph contrastive learning methods.
  4. In standard graph contrastive learning, the input views need to be changeable every epoch, for example randomly mask features or delete edges. GraphALFA tooks unchanged input views.
3 comments

Thank you for the detailed, constructive feedback!

W1: While the individual feature construction and contrastive learning components build upon established techniques, GraphAlFA's main contribution is the propagation-aware alignment mechanism, which conditions feature augmentation on the downstream GNN's observed spectral propagation behavior. This enables the augmentation to adapt to the graph, task supervision, and backbone. The broader feature-structure-label alignment formulation provides a unified framework for this mechanism. Importantly, we show that combining the standard components alone is not sufficient (Table 2), specifically, concatenating the standard attribute- and structure-derived views without the proposed alignment generally underperforms GCN on four of five datasets (“Concat” in Table 2), whereas the full framework consistently improves. Furthermore, removing either alignment objective degrades performance (Tables 2 and 8), supporting the contribution of the proposed alignment mechanism beyond combining standard components. We note that GraphAlFA is designed to be flexible with respect to the choice of structure-derived representation, rather than introducing a new structural embedding method. We employ established techniques as efficient, widely applicable implementations. We empirically evaluate various structural embedding (structure-view) configurations (Table 7) and demonstrate the framework's generalizability and effectiveness. LINE is selected as the default due to its low preprocessing cost (Table 14), contributing to the framework's overall computational efficiency (Table 3). More recent structural or positional encoders can be plugged in directly.

W2: We note that GraphAlFA computes the per-layer band-energy ratios only when it refreshes g_emp, every U epochs (Algorithm 1, lines 11-21). Between updates, the spectral alignment loss reuses the most recent profile. For each energy estimate, stochastic Lanczos spectral slicing uses a shared low-dimensional Krylov basis across bands and avoids a full Laplacian eigendecomposition. As analyzed in Appendix B.1.4, the dominant spectral computation scales linearly with graph size. Empirically, on Arxiv-Year with GCN, Table 9 reports that refreshing g_emp adds 14.80 ms per epoch (7.64% of training time) at U=10 and 4.86 ms per epoch (3.09%) at U=20, averaged over training. Table 3 shows that GraphAlFA achieves the lowest or near-lowest overall augmentation runtime among the compared methods, supporting the computational efficiency of our framework.

Q1: While both alignment objectives incorporate task supervision, they impose distinct constraints on the learned representations rather than performing the same feature-label alignment (defined in Sections 3.3.1 and 3.3.2). The augmented feature-label alignment loss shapes the geometry of node representations by promoting intra-class compactness and inter-class separation. In contrast, the augmented feature-structure alignment loss matches the relative spectral energy profiles of the augmented features and train-masked labels across graph Laplacian frequency bands, with each band's contribution weighted by the downstream GNN's empirically observed propagation behavior. Their complementary roles are empirically supported by our ablation studies (Tables 2 and 8), which support the benefit of combining these alignments, e.g., as shown in Table 2, on Chameleon with GCN, the full model achieves 48.69%, compared with 44.16% with only SupCon and 43.60% with only spectral alignment. We clarify this distinction of the structure- and propagation-conditioned nature of the spectral alignment in Section 3.3.

Q2: We note that feature-structure-label alignment is not an assumption about the data, but an optimization objective for addressing existing misalignment. Through task-supervised feature augmentation, the framework aims to construct representations that better exploit complementary attribute and structural information while accounting for the downstream GNN's propagation behavior. This alignment concerns the learned augmented features and their task relevance, rather than requiring agreement between graph connectivity and class labels or enforcing homophily. Our empirical results on heterophilic datasets support this capability under imperfect alignment. Gains over the original features are substantially larger on heterophilic than on homophilic graphs (Table 1; average +5.5 vs. +0.7 points with GCN and +7.4 vs. +2.9 with GAT). The controlled cSBM experiments (Figure 2) demonstrate improvements across varying feature and structural signal conditions, where gains for GCN and GAT are largest in heterophilic settings. Furthermore, the robustness analysis (Table 12) shows that GraphAlFA remains effective under feature and structural perturbations, with adaptive fusion adjusting the relative contributions of the two views. Appendix D.3.7 provides a complementary measure of how the augmented features relate to class labels, by showing that the adjusted feature homophily rises from 0.057 to 0.290 on Arxiv-Year and from 0.010 to 0.173 on Cora. Accuracy also improves on nodes whose neighborhoods are mostly dissimilar (Table 13: Arxiv-Year, local homophily ≤ 0.25: 39.6 → 45.4; Cora, (0.25, 0.5]: 46.2 → 55.3).

Q3: We agree that including graph contrastive learning methods would strengthen the empirical assessment. We have conducted additional experiments with representative methods (GRACE [1], GraphACL [2], and PolyGCL [3]). Specifically, for each contrastive learning baseline, we train its encoder using the recommended settings, then use the learned node embeddings with the same backbones. We have incorporated these additional baseline comparisons in table 1 of the manuscript. GraphAlFA achieves the best overall performance, with average improvements of 2.7 to 4.1 percentage points across all datasets and backbones. These results further demonstrate the effectiveness of our propagation-aware feature augmentation.

[1] Zhu et al. Deep Graph Contrastive Representation Learning (2020)

[2] Xiao, T., Zhu, H., Chen, Z., & Wang, S. Simple and Asymmetric Graph Contrastive Learning without Augmentations (NeurIPS 2023)

[3] Chen, J., Lei, R., & Wei, Z. PolyGCL: GRAPH CONTRASTIVE LEARNING via Learnable Spectral Polynomial Filters (ICLR 2024)

Q4: We note that while stochastic views are often constructed through masking or perturbations in self-supervised graph contrastive learning to form positive pairs, stochastic perturbations are not required [1]. GraphAlFA follows a different paradigm, where positive pairs are defined by shared task labels, following a supervised contrastive learning setup [2, 3]. This objective promotes class-discriminative representations without requiring independently perturbed contrastive views. Accordingly, the attribute-derived and structure-derived views capture complementary information, rather than stochastic contrastive views, and regenerating them at each epoch is not required. Although input views are precomputed (Alg. 1, lines 1–7; App. B.1.1), their learnable projections and adaptive fusion (Sec. 3.2) are updated under task supervision (Sec. 3.3; Alg. 1, lines 22–32), providing evolving augmented features. This allows GraphAlFA to learn task-relevant representations while preserving the observed graph topology. Stochastic view perturbations could be incorporated as an extension. We clarify this methodological distinction in Sections 3.1 and 3.3.

[1] Xiao, T., Zhu, H., Chen, Z., & Wang, S. Simple and Asymmetric Graph Contrastive Learning without Augmentations (NeurIPS 2023)

[2] Wang et al. ClusterSCL: Cluster-Aware Supervised Contrastive Learning on Graphs (WWW 2022)

[3] Khosla et al. Supervised Contrastive Learning (NeurIPS 2020)

Thank you again for your thoughtful and constructive feedback. We would like to check whether our initial response has adequately addressed your concerns, or if there are any remaining questions or points you would like us to clarify further. We would be happy to make additional revisions or provide further details if needed.

Reviewer_jr7DOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies feature-structure-label misalignment and offers theoretical insights. It uses a propagation-aware spectral objective to supervise alignment. However, the manuscript suffers from presentation flaws and lacks sufficient empirical support. More importantly, the authors omit a critical discussion: the proposed alignment process may lead to overfitting, as it essentially acts as a form of task-specific adaptation. Therefore, I lean toward a weak reject.

Critical strengths

None.

Critical weaknesses
  1. Many critical claims in this paper lack both literature and experimental support. e.g. The effectiveness of GNNs rests on an implicit data-level assumption: node attributes, graph structure, and task labels should induce mutually consistent notions of similarity.

Such misalignment can appear as noisy or weakly informative features, graph neighborhoods that connect task-dissimilar nodes, or spectral components that are poorly preserved by the downstream model.

Feature augmentation is particularly attractive because it enriches the downstream inputs without altering the observed topology or backbone architecture. However, existing methods primarily determine what structural, positional, or neighborhood information to inject or how to transform it, without explicitly accounting for how GNNs propagate the resulting signals. This creates an important disconnect: augmented information may be task-relevant at the input yet attenuated during downstream propagation, such that augmentation designed without accounting for the GNN can introduce signals that are subsequently suppressed.

.....

  1. The proposed alignment process may lead to overfitting, as it essentially acts as a form of task-specific adaptation. This is a critical concern for such studies. The authors should carefully analyze how the model mitigates this issue and discuss its practical limitations.

  2. The manuscript requires significant editing to improve readability, as many overly long sentences make it difficult for readers to follow.

Questions for authors
  1. Many critical claims in this paper lack both literature and experimental support. Where is the literature or empirical support for the following claims?

e.g. The effectiveness of GNNs rests on an implicit data-level assumption: node attributes, graph structure, and task labels should induce mutually consistent notions of similarity.

Such misalignment can appear as noisy or weakly informative features, graph neighborhoods that connect task-dissimilar nodes, or spectral components that are poorly preserved by the downstream model.

Feature augmentation is particularly attractive because it enriches the downstream inputs without altering the observed topology or backbone architecture. However, existing methods primarily determine what structural, positional, or neighborhood information to inject or how to transform it, without explicitly accounting for how GNNs propagate the resulting signals. This creates an important disconnect: augmented information may be task-relevant at the input yet attenuated during downstream propagation, such that augmentation designed without accounting for the GNN can introduce signals that are subsequently suppressed.

  1. The proposed alignment process may lead to overfitting, as it essentially functions as a task-specific adaptation. Given that this is a critical concern for such frameworks, the authors should clarify: a) How is this potential overfitting risk analyzed in this work? b) Is there any explicit mechanism within the proposed framework designed to mitigate this issue?
2 comments

Thank you for the detailed, constructive feedback!

W1+Q1: We have revised the Introduction section of the manuscript to strengthen the literature support and clarify the motivating arguments. The motivating observations are grounded in two established directions. First, prior studies demonstrate that GNN performance depends on the interaction between node attributes, graph structure, and task supervision, beyond conventional homophily assumptions [1,2], and that the usefulness of individual features can change under graph propagation [4]. Building on this perspective, we formulate feature augmentation as task-dependent feature-structure-label alignment, without requiring strict consistency or homophily. Second, spectral analyses show that GNN propagation can selectively attenuate or transform different frequency components in a backbone- and task-dependent manner [3]. This motivates incorporating the downstream backbone's observed spectral behavior into the augmentation process. Together, these observations suggest that effective feature augmentation should consider not only what information is introduced but also how it is processed by downstream propagation. Unlike representative feature augmentation approaches that primarily focus on constructing or transforming features, GraphAlFA incorporates empirical propagation behavior into its alignment objective. Empirically, our synthetic experiments (Figure 2) demonstrate improvements across varying feature and structural signal conditions, while the component ablations (Table 2) and local homophily analysis (Table 13) provide further evidence supporting the effectiveness of alignment-driven augmentation.

[1] Luan et al. When Do Graph Neural Networks Help with Node Classification? Investigating the Impact of Homophily Principle on Node Distinguishability (NeurIPS 2023)

[2] Zheng, Y., Luan, S., & Chen, L. What Is Missing In Homophily? Disentangling Graph Homophily For Graph Neural Networks (NeurIPS 2024)

[3] Dong et al. Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025)

[4] Navarro, M., Azizpour, A., & Segarra, S. Adaptive Node Feature Selection For Graph Neural Networks (ICML 2026)

W2+Q2: (a) We examine generalization from both theoretical and empirical perspectives. Appendix B.3 provides a transductive PAC-Bayes analysis that characterizes the relationship between augmentation complexity and generalization risk. Empirically (in Tables 1,10,11,12), our extensive evaluation across diverse datasets and backbone architectures demonstrates consistent improvements on held-out test sets. Specifically, at 1% labels, Table 11 reports 37.30% versus 32.63% for Chameleon and 35.84% versus 32.90% for Squirrel against base GCN. That shows the gains persist in those low-label settings. (b) To mitigate overfitting, GraphAlFA leverages residual feature augmentation, which preserves the original input as a recoverable special case, together with regularization λreg∥ΔXθ∥F2\lambda_{\mathrm{reg}}\|\Delta X_\theta\|_F^2 that discourages excessive modifications to the original features (Section 3.2). This design is also motivated by the augmentation-complexity perspective in Appendix B.3.

W3: Thank you for the feedback! We have revised the manuscript to improve readability, particularly by shortening overly long sentences and clarifying complex explanations, including the framework overview in introduction, spectral alignment formulation in Section 3.3, and ablation discussion in Section 5.2.

Thank you again for your thoughtful and constructive feedback. We would like to check whether our initial response has adequately addressed your concerns, or if there are any remaining questions or points you would like us to clarify further. We would be happy to make additional revisions or provide further details if needed.

Reviewer_R4cQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GraphALFA, a supervised feature augmentation framework for node classification that combines multi-hop attribute context with structural embeddings and aligns the augmented representation with labels and the downstream GNN’s estimated propagation behavior. Its propagation-aware component weights spectral feature-label alignment using an empirical band-wise energy response measured across GNN layers. The method is evaluated on multiple real-world datasets and synthetic cSBM graphs.

Critical strengths
  1. The motivation is relevant and clearly presented. Adapting feature augmentation to the behavior of the downstream GNN is a reasonable direction, particularly when feature, structure, and label signals are not naturally aligned.

  2. The empirical evaluation is broad.

Critical weaknesses
  1. I am not convinced that the proposed �0��0� can be interpreted as the downstream GNN’s frequency response for the general backbones considered. The ratio of output to input energy in a frequency band measures net energy change, but for nonlinear or attention-based GNNs, energy can move across bands rather than being independently attenuated or preserved. The theoretical justification, in contrast, assumes a diagonal spectral-filter approximation �1��1�, under which such cross-band mixing does not occur. This leaves a substantial gap between the central propagation-aware interpretation and the models used in the experiments.

  2. The novelty and empirical positioning are not sufficiently convincing for an ICLR contribution. Most components—multi-hop propagated features, off-the-shelf structural embeddings, gated fusion, supervised contrastive learning, and spectral energy matching—are established techniques, with the empirical-response weighting appearing to be the main new element. At the same time, the experimental comparison is relatively narrow and dated: the five feature-augmentation baselines range from 2022 to 2025, and directly relevant work is omitted, with no 2026 augmentation methods included despite the ICLR 2027 timing.

Questions for authors
  1. How does �0��0� distinguish preservation of energy already present in a frequency band from energy transferred into that band by nonlinearities, attention, or other cross-frequency interactions? A band-to-band transfer analysis, rather than a scalar energy ratio, seems necessary to support the stated frequency-response interpretation.

  2. The “attribute-derived” view is �1��1�, and therefore already contains substantial topology-dependent information. How should the claimed complementarity between attribute- and structure-derived views be interpreted, and how can the contribution of structural information be isolated when the “No struct view” variant still retains �2�and�3��2� and �3�?

  3. Why does the baseline suite omit directly relevant recent node feature augmentation methods such as LGGD [1], and why are there no 2026 augmentation baselines? Given that the paper describes the comparison set as state of the art, a more current comparison seems important for assessing the empirical contribution.

[1] A Learned Generalized Geodesic Distance Function-Based Approach for Node Feature Augmentation on Graphs

3 comments

Thank you for the detailed, constructive feedback!

W1+Q1: We agree that for nonlinear or attention-based GNNs, the band-wise energy ratio g_emp reflects the net change in band-wise spectral energy, which doesn't distinguish between within-band preservation and cross-band energy transfer. This is consistent with [1], which shows that non-linear layers can substantially alter the frequency distribution of GNN outputs, which is why we treat g_emp as an observed band-energy ratio rather than an exact filter response (Sec. 3.3.2). Accordingly, it should be interpreted as an empirical characterization of the backbone's observed spectral energy behavior instead of an exact diagonal frequency response. Our theoretical analysis in Section 4.2 establishes the relationship between these energy ratios and spectral transmission under a diagonal-filter approximation, in a restricted linear-filter setting, under which band-wise energy ratios admit a direct interpretation in terms of spectral transmission. In practice, we empirically measure the band-wise spectral energy changes, incorporating the effects of nonlinearities and attention. Importantly, the propagation-aware objective uses these empirical measurements to guide spectral alignment, without requiring an explicit analytical filter. We have revised Sections 3.3 and 4.2 in the manuscript to clarify this distinction, describing g_emp as an observed band-energy ratio and restricting the spectral transmission interpretation to the diagonal linear-filter approximation.

[1] Dong et al. Graph Neural Networks Are More Than Filters: Revisiting and Benchmarking from A Spectral Perspective (ICLR 2025)

W2+Q3: While multi-hop feature propagation, structural embeddings, and contrastive learning components build upon established techniques, the central contribution of GraphAlFA is a propagation-aware feature augmentation framework that conditions task-supervised spectral alignment on the downstream backbone's empirically observed propagation behavior. This adapts the augmented features to the graph, supervision, and backbone jointly. The broader feature-structure-label alignment formulation provides a unified framework for this mechanism. Unlike conventional feature augmentation approaches that primarily focus on constructing additional features, GraphAlFA additionally accounts for how the downstream GNN processes that information. This enables feature augmentation that adapts jointly to the graph, supervision, and backbone, without modifying the graph topology or model architecture. The component ablations (Tables 2 and 8) further demonstrate that simple feature concatenation or removing alignment objectives consistently leads to inferior performance, supporting the importance of the proposed integrated design. We agree that incorporating more recent feature augmentation baselines would strengthen the empirical assessment. We have conducted additional experiments with three relevant methods (LGGD [1], SCOUT [2], and FAF [3]). The results are added to table 1 in the manuscript. GraphAlFA outperforms these additional baselines across settings, by 4.3 to 5.2 percentage points on average. These results further demonstrate the effectiveness of our propagation-aware alignment. We have incorporated these additional baselines and updated the baseline discussions.

[1] Azad, A., & Fang, Y. A Learned Generalized Geodesic Distance Function-Based Approach for Node Feature Augmentation on Graphs (KDD 2024)

[2] Seo et al. SCOUT: Structure-Aware Aspect and Anchor-Count Selection for Node Attribute Augmentation via Positional Information (WWW 2026)

[3] Rubio-Madrigal, C., & Burkholz, R. Fixed Aggregation Features Can Rival GNNs (ICML 2026)

Q2: Both views incorporate structural information, but through distinct mechanisms and thus capture complementary forms of information. The attribute-derived view (X_f) represents topology-conditioned attributes, whereas (X_s) encodes attribute-independent structural embeddings. The complementarity is therefore between attribute-conditioned neighborhood information and attribute-independent structural information. Accordingly, the “No struct view” ablation (Table 2) specifically refers to removing the structure-derived embedding view (X_s), rather than eliminating all topological information. It retains (X_f) unchanged, thereby evaluating the additional contribution of topology-only structural embeddings beyond attribute propagation. A fully topology-free variant would not be a valid ablation in our setting, since the GNN backbone itself propagates over the graph, and the spectral alignment objective is defined using the graph Laplacian. Instead, the “No struct view” ablation isolates the incremental contribution of (X_s), while the “Only Ego Feat.” ablation removes propagated attributes but retains (X_s), and with the MLP backbone (Table 8) structure enters only through the augmentation. The complementary contribution is also noted in Table 12, which reports the average gate weight on the attribute-derived view X_f for the Chameleon dataset: With 20% feature masking, this weight drops from 0.571 to 0.534, shifting weight toward the structure-derived view X_s; with 10% edge dropping, it rises to 0.634, shifting weight toward X_f. This indicates that GraphAlFA treats them as distinct sources and adjusts how it combines the two views. We clarify the intended interpretation of the two views and their distinctions in Sections 3.1 and 5.2.

Thanks for the response. The clarifications and the additional baselines address several of my concerns. However, I still feel that most components were established and the contribution is incremental. So I keep my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.