TL;DR
This paper addresses the challenge of missing node features in GNNs, proposing new datasets, evaluation protocols, theoretical analysis, and a robust baseline method called GNNmim for improved node classification.
Contribution
It introduces realistic datasets and evaluation protocols for missing features, provides theoretical insights into missingness mechanisms, and proposes a robust baseline method GNNmim.
Findings
High sparsity limits information loss, making models seem robust.
GNNmim performs competitively across datasets and missingness scenarios.
New datasets and evaluation protocols better reflect real-world missing data challenges.
Abstract
Handling missing node features is a key challenge for deploying Graph Neural Networks (GNNs) in real-world domains such as healthcare and sensor networks. Existing studies mostly address relatively benign scenarios, namely benchmark datasets with (a) high-dimensional but sparse node features and (b) incomplete data generated under Missing Completely At Random (MCAR) mechanisms. For (a), we theoretically prove that high sparsity substantially limits the information loss caused by missingness, making all models appear robust and preventing a meaningful comparison of their performance. To overcome this limitation, we introduce one synthetic and three real-world datasets with dense, semantically meaningful features. For (b), we move beyond MCAR and design evaluation protocols with more realistic missingness mechanisms. Moreover, we provide a theoretical background to state explicit…
Peer Reviews
Decision·Submitted to ICLR 2026
S1: The main premise is quite interesting, and having myself evaluated GNN models on sparse datasets, I can clearly see how it could have been flawed. The experiments and theoretical analysis (sparsity and information loss connection) make argument intuitive. S2: Contributes datasets that elicits the flaws of earlier benchmarks by showing that even a simple GNN (with mask for missing features) is competitive. S3. Paper is well written, easy to follow, and claims are supported by correspondi
W1. The graph scale of real-world dataset is quite small. Most real datasets currently have millions of nodes. How do results generalize to large scale graphs with dense features? Also will GNNmim remain competitive in such graphs? Further the benchmark consists of 3 graphs – they may not be representative of real graphs. How non-trivial it is to expand the benchmark? W2. The paper emphasizes on GNNmim as a new method, however, the main contribution seems to be the benchmark developed to assess
1. The motivation to investigate why existing benchmarks and missingness assumptions fail to reflect real-world scenarios is well justified. 2. The proposed GNNmim framework is simple yet effective, showing consistent robustness across various missing-feature settings.
1. The information-theoretic results (e.g., ∆ bounds/data-processing) are important but scattered; adding “takeaway” remarks after each theorem and clarifying practical conditions (e.g., bandwidth/feature distributions) would help adoption. 2. More direct comparisons and discussion versus imputation-free conditional models and density-oriented approaches are needed. 3. The experimental studies are not strong enough since they are conducted on limited benchmark datasets and lack evaluations on la
1.Theoretical proof: The paper reveals the limitations of traditional sparse datasets when evaluating the robustness of missing features, making the evaluation process more scientifically grounded. 2.Introduction of new datasets: New datasets with dense, raw, and semantically meaningful features are introduced, providing more challenging and representative benchmarks for future research. 3.Proposing a more robust model: GNNmim is a simple yet effective new approach that consistently outperform
1.The proposed synthetic/real-world datasets have relatively few features, reflecting the nature of the original measurements, but this also limits the direct generalizability of the findings in high-dimensional sparse feature scenarios. 2.GNNmim uses zero-padding as a placeholder for missing values, which is an arbitrary replacement rather than a semantically meaningful feature attribution. More theoretical analysis is needed to justify the appropriateness of this approach. Additionally, the e
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Machine Learning in Healthcare · Big Data and Digital Economy
