Less is More: Towards Simple Graph Contrastive Learning
Pith reviewed 2026-05-18 11:20 UTC · model grok-4.3
The pith
A simple graph contrastive learning model without data augmentation or negative sampling achieves state-of-the-art results on heterophilic graphs by treating original node features and graph structure as complementary views.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The work establishes that mitigating node feature noise by aggregating it with structural features derived from the graph topology allows the original node features and the graph structure to function as two complementary views for contrastive learning. An embarrassingly simple model using a GCN encoder to capture structural features and an MLP encoder to isolate node feature noise therefore suffices for effective unsupervised representation learning without data augmentation or negative sampling, attaining state-of-the-art results on heterophilic benchmarks along with advantages in complexity, scalability, and robustness.
What carries the argument
Dual-encoder setup with a GCN capturing structural features from graph topology and an MLP isolating node feature information, enabling contrastive learning from the graph's inherent complementary views.
Load-bearing premise
The original node features and the graph structure naturally provide two complementary views for contrastive learning such that their combination suffices for effective unsupervised representation learning.
What would settle it
Experiments on additional heterophilic graph datasets where the proposed model fails to match or exceed the accuracy of methods that rely on augmentation and negative sampling would undermine the claim that such complexity is unnecessary.
read the original abstract
Graph Contrastive Learning (GCL) has shown strong promise for unsupervised graph representation learning, yet its effectiveness on heterophilic graphs, where connected nodes often belong to different classes, remains limited. Most existing methods rely on complex augmentation schemes, intricate encoders, or negative sampling, which raises the question of whether such complexity is truly necessary in this challenging setting. In this work, we revisit the foundations of supervised and unsupervised learning on graphs and uncover a simple yet effective principle for GCL: mitigating node feature noise by aggregating it with structural features derived from the graph topology. This observation suggests that the original node features and the graph structure naturally provide two complementary views for contrastive learning. Building on this insight, we propose an embarrassingly simple GCL model that uses a GCN encoder to capture structural features and an MLP encoder to isolate node feature noise. Our design requires neither data augmentation nor negative sampling, yet achieves state-of-the-art results on heterophilic benchmarks with minimal computational and memory overhead, while also offering advantages in homophilic graphs in terms of complexity, scalability, and robustness. We provide theoretical justification for our approach and validate its effectiveness through extensive experiments, including robustness evaluations against both black-box and white-box adversarial attacks.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims that a simple graph contrastive learning approach, using a GCN encoder to capture structural features from the graph topology and an MLP encoder to isolate node feature noise, provides two complementary views without requiring data augmentation or negative sampling. This design is said to achieve state-of-the-art results on heterophilic benchmarks with minimal overhead, offer advantages in homophilic graphs regarding complexity and robustness, and be supported by theoretical justification plus extensive experiments including adversarial robustness tests.
Significance. If the empirical results and theoretical claims hold upon full inspection, the work would be significant for simplifying GCL methods, especially in heterophilic settings where existing approaches struggle. Demonstrating that complex augmentations and negative sampling are unnecessary could reduce computational barriers and encourage more scalable unsupervised graph learning techniques.
major comments (2)
- Abstract: the central claim of SOTA performance on heterophilic benchmarks and theoretical justification is asserted without any specific quantitative results, baselines, dataset details, derivation steps, or experiment descriptions, which are load-bearing for evaluating the soundness of the proposed principle and its advantages over existing GCL methods.
- Abstract: the observation that 'original node features and the graph structure naturally provide two complementary views' is presented as the foundation for the GCN+MLP design, but lacks any supporting sketch, equation, or analysis to show why aggregating feature noise with structural features suffices for effective contrastive learning.
Simulated Author's Rebuttal
We thank the referee for the constructive feedback on our manuscript. We address the major comments point by point below. We agree that the abstract can be improved for clarity and have revised it to incorporate additional supporting details from the main text while preserving its conciseness.
read point-by-point responses
-
Referee: Abstract: the central claim of SOTA performance on heterophilic benchmarks and theoretical justification is asserted without any specific quantitative results, baselines, dataset details, derivation steps, or experiment descriptions, which are load-bearing for evaluating the soundness of the proposed principle and its advantages over existing GCL methods.
Authors: We acknowledge that the abstract presents a high-level summary of the claims. The specific quantitative results, baselines, dataset details, and experiment descriptions are provided in Sections 4 and 5, while the theoretical justification and derivation steps appear in Section 3. To address the concern, we have revised the abstract to include concise quantitative highlights of the SOTA performance on heterophilic benchmarks and a brief reference to the theoretical analysis. revision: yes
-
Referee: Abstract: the observation that 'original node features and the graph structure naturally provide two complementary views' is presented as the foundation for the GCN+MLP design, but lacks any supporting sketch, equation, or analysis to show why aggregating feature noise with structural features suffices for effective contrastive learning.
Authors: The observation follows from our analysis that node features in heterophilic graphs often contain noise that can be mitigated by aggregation with structural features from the topology, yielding complementary views (one via GCN for structure, one via MLP for features). This principle is elaborated with supporting analysis in the introduction and theoretical sections. We have revised the abstract to include a short explanatory clause referencing this noise-mitigation insight. revision: yes
Circularity Check
No significant circularity
full rationale
With only the abstract available, the paper presents an empirical observation about node features and graph structure as complementary views, leading to a simple GCN+MLP design without augmentation or negative sampling. No equations, fitting procedures, or derivation steps are shown that reduce a claimed prediction or first-principles result to its inputs by construction. The theoretical justification is asserted but not detailed, and no self-citations or ansatzes are quoted that could create a load-bearing circular chain. The central claims rest on experimental validation rather than any self-referential reduction, making the derivation self-contained against the provided text.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Node features and graph topology provide naturally complementary views that mitigate feature noise when aggregated for contrastive learning
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.