acceptodds
Withdrawn from ICLR 2027

MDFS-GNN: Multi-Dilation Graph Propagation with Feature Shrinkage for Semi-Supervised Node Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph neural networks (GNNs) have been widely used for semi-supervised node classification by propagating structural and feature information across graph neighborhoods. However, when only a limited number of labeled nodes are available, insufficient supervision restricts effective information propagation between labeled and unlabeled nodes. Although high-order GNNs can enlarge the receptive field to capture long-range structural information, their insufficient modeling of local information and ineffective integration of local and global features may limit the learning of discriminative node representations. To address these challenges, we propose MDFS-GNN, a Multi-Dilation Graph Propagation with Feature Shrinkage framework for semi-supervised node classification. Specifically, MDFS-GNN employs graph convolution operations with different dilation factors to capture multi-order neighborhood information and incorporates the original node attributes through residual aggregation to preserve informative features during propagation. Furthermore, a feature shrinkage mechanism reconstructs the neighborhood structure according to node feature similarity and retains the Top- most similar nodes, thereby strengthening informative local interactions while reducing redundant information. These components are integrated into a two-stage learning framework consisting of initialization and iterative refinement, in which node representations and pseudo-labels are progressively updated to facilitate information propagation across the graph. Experiments on three homophilic and four heterophilic graph datasets under different labeling rates demonstrate the effectiveness and robustness of MDFS-GNN, particularly in scenarios with extremely limited labeled nodes.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Missing Data, few-Label Learning, graph convolutional networks, graph representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_YJ2pOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper focuses on the problem of semi-supervised node classification under limited labeled nodes. The authors indicate that existing GNNs suffer from insufficient supervision and high-order GNNs limit the learning of discriminative node representations due to insufficient local modeling and ineffective integration of local and global features. To this end, they propose MDFS-GNN. The proposed method captures multi-order neighborhood information with different dilation factors, and reconstructs the neighborhood structure by retaining the most similar nodes. The experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The authors propose a multi-dilation graph propagation mechanism to capture both local patterns and long-range structural dependencies without increasing propagation depth.

  2. The authors develop a feature shrinkage strategy that leverages node feature similarity to refine graph structures, reducing the influence of redundant and noisy connections.

Critical weaknesses
  1. The novelty of the proposed method is limited. The problems of label sparsity and fixed receptive fields have both been explored. The authors combine them but do not provide a reasonable motivation.

  2. The illustration of the proposed method (Figure 2) is very blurry and appears to be AI-generated.

  3. The experimental analysis is too cursory. For example, in Section 3.2, lines 308 to 315 are used to introduce the comparison methods, and only one sentence, "As shown in Table 3, MDFS-GNN achieves competitive or superior performance compared with existing methods" is used for the result analysis.

  4. The authors state that they report the average accuracy over 10 runs. However, the standard deviation is not provided in the experimental results.

  5. There are some inconsistent statements in the experiment. For example, line 417 should be "Table 5 compares the base model and MDFS-GNN," and line 410 should be "Figure 2 presents the classification performance under different labeling rates.".

  6. The comparison methods in this manuscript are unreasonable. The performance of the proposed method is attributed to the use of pseudo-labels, yet the pseudo-label-based methods compared by the authors are limited.

Questions for authors
  1. What is the motivation for jointly addressing the label sparsity problem and the fixed receptive field problem?
Reviewer_ygFhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses semi-supervised node classification under very low label rates. It proposes MDFS-GNN, which combines (i) a multi-dilation propagation module that aggregates adjacency powers at strided orders (Eq. 2 to 4) with residual concatenation of initial features, (ii) a feature shrinkage module that rebuilds the adjacency by keeping the top-K most cosine-similar nodes in the initial feature space (Eq. 6 to 8), and (iii) a two-stage scheme where a logistic regression classifier is trained with pseudo-labels that are iteratively updated while ground-truth labels are reset each round (Eq. 9 to 13). Experiments on three homophilic and four heterophilic datasets report gains concentrated in the extreme low-label regime, for example +7.2 and +7.6 points over the best baseline on Cora and CiteSeer with one label per class (Table 4).

Critical strengths
  1. The low-label improvements are sizable and internally consistent. Table 4 shows gains of 7.2 to 10.5 points at one to two labels per class, and the ablation in Table 6 shows feature shrinkage contributes most exactly in that regime (55.7 to 74.5 on Cora with one label), so the central claim is supported by more than one cut of the data.
  2. overall, the paper is well-presented and easy to follow
Critical weaknesses
  1. The conceptual delta over prior work is small and not clearly articulated. Multi-dilation propagation is a strided subset of adjacency powers closely related to MixHop and polynomial propagation, feature shrinkage is feature-similarity kNN graph reconstruction as in graph structure learning, and the pseudo-label loop is standard self-training. The paper does not analyze what is fundamentally new relative to GraphHop, the closest two-stage propagation method it cites and compares against.
  2. The evidence is narrow and partly unfavorable to the method. On standard splits (Table 3) MDFS-GNN trails GCNII on Cora (81.9 vs 85.7) and PubMed (78.0 vs 79.8), the heterophilic results appear only as a figure without tabulated numbers, all datasets are small, and the ablation covers only the feature shrinkage module, leaving the contributions of multi-dilation propagation and the two-stage strategy unquantified. The strong dependence on the per-dataset tuned K (Figure 4) further limits the practical takeaway.
Questions for authors

Refer to weaknesses.

Reviewer_H6p6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes MDFS-GNN, a semi-supervised GNN model designed for node classification under limited label supervision. The framework combines multi-dilation graph propagation, feature similarity-based graph reconstruction, and iterative pseudo-label refinement. Specifically, multi-dilation propagation aggregates neighborhood information across different receptive fields, while feature shrinkage constructs a Top-K similarity graph to strengthen informative node interactions. A two-stage learning strategy further updates node representations and pseudo-labels using a logistic regression classifier. Experiments on three homophilic and four heterophilic benchmark datasets suggest that MDFS-GNN achieves competitive performance, particularly in extremely low-label settings.

Critical strengths

1 This paper addresses a practically important problem in graph representation learning, where the technique of this paper is motivated by the challenge of limited labelling signal in node classification. 2 The proposed method exhibits promising performance even under extremely sparse supervision.

Critical weaknesses

1 Limited novelty. The proposed model of this paper mainly combines existing graph learning techniques. This paper mainly propose three modules: multi-dilation graph propagation mechanism which introduces different multi-order adjacency matrices for propagation; feature shrinking module which establishes a new adjacency matrix using simply feature similarity; iterative pseudo-label refinement strayegy which optimizes the model using pseudo labels. These components closely resemble existing approaches, such as MixHop and GraphHop. The paper does not clearly establish the fundamental methodological differences or demonstrate how their integration provides substantial advances beyond existing methods. Besides, The paper provides no theoretical or empirical evidence to support the claim that the proposed module can eliminate irrelevant nodes, redundant information, and noise. These challenges, however, are what motivated the authors to write this paper (see lines 59-64). 2 Ambiguous methodological formulation and insufficient experimental validation. Several key aspects of the proposed framework are inadequately specified. In particular, the X\mathbf{X} in Eq. (8) is ambiguoous, since it is inconsistent with the part of Eq. (7); Eq. (9) does not clarify how the logistic regression classifier jointly updates node representations and pseudo-labels, while Eq. (8) uses identical notation for the input and updated features. Furthermore, the ablation studies is insufficient without independently examining multi-dilation propagation and iterative refinement. The inconsistent results across the main comparison and ablation tables also remain unexplained.

Questions for authors

Except the questions proposed in Critical Weaknesses, there are still some additional questions the authors should be addressed. (1) Lack of computational complexity. These modules require additional overhead, potentially introducing quadratic computational and memory overhead. This paper does not appear to offer any advantages over other baselines in terms of complexity. (2) Insufficient hyperparameter analysis. The experiments mainly investigate the influence of K, while other important hyperparameters, such as the number of dilation orders, dilation intervals, and pseudo-label loss weight, are not systematically examined. How sensitive is the proposed method to these parameters? (3) Experimental reliability. Although the experiments are repeated ten times, most comparison tables report only average accuracy without standard deviations. Can the authors provide variance estimates or statistical significance tests to establish the reliability of the reported improvements? (4) Presentation and notation issues. Several equations and references require clarification. For example, Eq. (8) uses the same notation for features before and after updating, and some figure and table references are inconsistent with their actual numbering. The authors should carefully revise these issues to improve the clarity and reproducibility of the paper. (5) Lack of theoretical analysis. The paper does not provide sufficient theoretical justification for the proposed modules for supporting the advantage of MDFS-GNN, or effectively addressing the problem inherent in the existing related algorithms. Besides, can the authors explain why their combination improves information propagation and alleviates over-smoothing, particularly under extremely limited supervision?

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

Related papers

Discussion (0)

Sign in to comment.