Research map / Graph learning / GNN architectures
Rethinking GNNs and Missing Features: Challenges, Evaluation and a Robust Solution
Francesco Ferrini, Veronica Lachi, Antonio Longa, Bruno Lepri, Akiyoshi Matono, Andrea Passerini, Xin Liu, Manfred Jaeger
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| concatenate node features with binary missingness mask (MIM) in RDL | RDL · REL-TRIAL site-success | MAE: 0.4235 -> 0.3746 | Improved |
| add MIM mask to RDL | RDL · REL-EVENT user-attendance | MAE: 0.2628 -> 0.2553 | Improved |
| add MIM mask to RDL | RDL · REL-TRIAL study-adverse | MAE: 43.5251 -> 43.1148 | Improved |
| add MIM mask to RDL | RDL · REL-F1 driver-top3 | ROC-AUC: 0.755 -> 0.768 | Improved |
| add MIM mask to RDL on nearly complete data | RDL · REL-ARXIV author-publication (0.07% missing) | MAE: 0.5033 -> 0.5051 | No clear effect |
| add MIM mask to RDL | RDL · RelBench runtime | runtime per seed: 1512.0 -> 1526.2 s (REL-EVENT) | No clear effect |
| evaluate on dense low-dimensional datasets instead of sparse bag-of-words benchmarks | existing GNN missing-feature methods · Cora, Citeseer, PubMed vs Synthetic, Air, Electric, Tadpole | F1 vs missingness | Hurt |
| increase missingness rate on sparse benchmarks | all GNN methods · Cora, Citeseer, PubMed | F1 | No clear effect |
| switch missingness mechanism from uniform to feature-dependent MNAR | GOODIE, FairAC · Synthetic / Electric | AUC of F1: GOODIE 0.771 -> 0.587 (CD-MNAR); FairAC 0.870 (Electric S-MCAR) vs 0.641 (Synthetic FD-MNAR) | Hurt |
| train on FD-MNAR 50% and test on U-MCAR (distribution shift in missingness) | GNN missing-feature methods · R2 regime | F1 | Hurt |
| use GNNmim versus other missing-feature GNNs | GNN with missing node features · all missingness mechanisms and datasets | AUC of F1 | Improved |
| use GNNmim versus MICE, MLP+MIM and XGBoost under CD-MNAR | GNNmim · CD-MNAR (Appendix L) | performance | Improved |
Closest papers by meaning
- T2-GNN: Graph Neural Networks for Graphs with Incomplete Features and Structure via Teacher-Student Distillation (AAAI 2023 · similarity 0.89)
- Propagate and Inject: Revisiting Propagation-Based Feature Imputation for Graphs with Partially Observed Features (ICML 2025 · similarity 0.87)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.86)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.86)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.86)
- Generalization of Graph Neural Networks Is Robust to Model Mismatch (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
