Research map / Graph learning / Graph self-supervised learning
DeepITE: Designing Variational Graph Autoencoders for Intervention Target Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeepITE, a variational graph autoencoder with GAT-based inference and a graph-surgery generative decoder that estimates intervention targets across graphs in self- or semi-supervised fashion.
Problem
Intervention target estimation (ITE) methods learn each intervention instance independently, cannot use labeled intervention data, and must be re-run when data or causal graphs change.
Approach
A VGAE whose generative model applies graph surgery via a Bernoulli intervention indicator per node on a non-linear GNN decoder, and whose GAT-based inference model amortizes posterior inference over intervention indicators, exogenous noise and noise precision, trained by ELBO maximization with optional supervised loss on labeled targets.
Key results
- On the synthetic Graph-50 dataset, DeepITE (sep) reaches Recall@1 0.723 and Recall@5 0.972, above the best baseline CauseInfer/CI-RCA-level results (e.g., CI-RCA Recall@5 0.952).
- On Protein Signaling, DeepITE achieves Recall@1 0.652 vs. 0.615 for the best baseline (ShapleyFlow); on ICASSP-SPGC 2022 it beats the best baseline by over 20% in Recall@1.
- Adding 5-10% labeled data improves DeepITE Recall@1 by roughly 4-20% depending on the dataset (e.g., Protein Signaling 0.652 to 0.842 at 5%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeepITE (sep) rather than the strongest baseline | ITE on synthetic graphs · Synthetic Graph-50, Graph-100, and Graph-500 | Recall@1: 0.647 -> 0.723; 0.631 -> 0.685; 0.623 -> 0.642 | Helped |
| Use DeepITE (sep) rather than the strongest baseline | ITE on synthetic graphs · Synthetic Graph-50, Graph-100, and Graph-500 | Recall@5: 0.952 -> 0.972; 0.930 -> 0.968; 0.823 -> 0.891 | Helped |
| Use DeepITE rather than the strongest baseline | ITE on Protein Signaling · Protein Signaling | Recall@1: 0.615 -> 0.652 | Helped |
| Use DeepITE rather than the strongest baseline | ITE on ICASSP-SPGC 2022 · ICASSP-SPGC 2022 | Recall@1: 0.677 -> 0.881 | Helped |
| Add 5% labeled data (ablation) | DeepITE · HeterITE-50, HeterITE-100, and HeterITE-500 | Recall@1: 0.718 -> 0.820; 0.690 -> 0.731; 0.627 -> 0.667 | Helped |
| Add 5% labeled data (ablation) | DeepITE · Protein Signaling | Recall@1: 0.652 -> 0.842 | Helped |
| Add 5% labeled data (ablation) | DeepITE · ICASSP-SPGC 2022 | Recall@1: 0.881 -> 0.906 | Helped |
| Replace the DeepITE decoder with the VACA decoder (ablation) | DeepITE · Graph-50 | MMD(Obs): 0.12 ± 0.075 -> 1.93 ± 0.38 | Hurt |
| Replace the DeepITE encoder with the DAG-GNN encoder (ablation) | DeepITE · Graph-50 | MMD(Obs): 0.12 ± 0.075 -> 2.58 ± 0.40 | Hurt |
| Replace the DeepITE decoder with the VACA decoder (ablation) | DeepITE · Graph-50 | MMD(Int): 0.76 ± 0.051 -> 10.78 ± 2.99 | Hurt |
| Replace the DeepITE encoder with the DAG-GNN encoder (ablation) | DeepITE · Graph-50 | MMD(Int): 0.76 ± 0.051 -> 32.80 ± 7.89 | Hurt |
| Replace the DeepITE decoder with the VACA decoder (ablation) | DeepITE · Graph-50 | SSE: 415.66 ± 26.92 -> 2534.70 ± 105.22 | Hurt |
Surprising findings
- DeepITE (mix), trained on graphs of all sizes together, outperforms DeepITE (sep) trained only on 100-node graphs in Recall@1.
- Traditional ITE methods like UT-IGSP and CITE perform much better on small graphs with large sample sizes, but fall short on larger graphs with fewer samples.
Limitations
- Assumes causal sufficiency and a known, acyclic causal graph adjacency matrix.
- Recall@1 declines as graph size increases.
- Requires labeled intervention targets for the semi-supervised gains.
Experimental setup
Datasets: Synthetic time-series causal graphs (50/100/500 nodes) Protein Signaling (Sachs et al.) ICASSP-SPGC 2022 5G KPI dataset VACA synthetic SCM data
Benchmarks: Synthetic (Graph-50/100/500) Protein Signaling ICASSP-SPGC 2022 Synthetic classical ITE setting (Linear-10/20, Nonlinear-10/20)
Metrics: Recall@1 Recall@5 Recall@3 root-cause accuracy root-cause score MMD SSE runtime
Baselines: UT-IGSP CITE PreDITEr TreeExplainer ASV ShapleyFlow PWSHAP CauseInfer MicroHECL MicroRCA CausalRCA CI-RCA
Builds on: DAG-GNN VACA Gumbel-softmax CI-RCA
Closest papers by meaning
- Towards Causal Foundation Model: on Duality between Optimal Balancing and Attention (ICML 2024 · similarity 0.85)
- Sample Efficient Bayesian Learning of Causal Graphs from Interventions (NeurIPS 2024 · similarity 0.84)
- Linear Causal Representation Learning from Unknown Multi-node Interventions (NeurIPS 2024 · similarity 0.84)
- Estimating Interventional Distributions with Uncertain Causal Graphs through Meta-Learning (NeurIPS 2025 · similarity 0.83)
- Causal Inference with Conditional Front-Door Adjustment and Identifiable Variational Autoencoder (ICLR 2024 · similarity 0.83)
- Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.83)
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.