Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Causally Invariant Feature Learning
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 FedCIFL, a federated feature selection method that iteratively estimates federated causal effects with sample reweighting and a supervised autoencoder to select causally invariant features under Non-IID and OOD settings.
Problem
Federated feature selection under Non-IID client data and out-of-distribution test data, where existing FFS methods rely on feature-label correlations that are affected by selection bias and fail to generalize.
Approach
Learns causally invariant features in the original feature space by reweighting samples to remove confounding bias, estimating per-feature causal effects with other features as confounders, and iteratively shrinking the confounder set; irrelevant feature subsets are aligned on the server via a privacy-preserving rank-based voting that does not require client sample sizes.
Key results
- Abstract reports beating the best-performing baseline by 3.19%, 9.07% and 2.65% in average test Accuracy, RMSE and F1 score, respectively.
- On Amazon Review with MLP, FedCIFL reaches 84.01% accuracy on BDE→K and 83.61% on BDK→E, outperforming all baselines on most tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedCIFL instead of the best-performing baseline | Federated feature selection · Synthetic and real-world datasets | average test Accuracy: +3.19% | Helped |
| Use FedCIFL instead of the best-performing baseline | Federated feature selection · Synthetic and real-world datasets | average test RMSE: +9.07% | Helped |
| Use FedCIFL instead of the best-performing baseline | Federated feature selection · Synthetic and real-world datasets | average test F1 score: +2.65% | Helped |
| Add iterative refinement of the confounder set (ablation) | FedCIFL · Synthetic Non-IID+OOD | Accuracy | Helped |
| Add the supervised autoencoder for low-dimensional representations (ablation) | FedCIFL · Synthetic Non-IID+OOD | Accuracy | Helped |
| Add weighted voting to resolve conflicts between multiple modes (ablation) | FedCIFL · Synthetic Non-IID+OOD | Accuracy | Helped |
Surprising findings
- On the Amazon Review DEK→B task, FedCIFL (73.05%) is slightly below the best baseline EAMB-V3 (73.40%).
Limitations
- Only binary-feature synthetic setup and a single real-world sentiment dataset are evaluated.
- Currently limited to binary classification; multi-label extension left for future work.
- Relies on assumptions (overlap, Assumption 1 invariance) for theoretical correctness.
Experimental setup
Datasets: Synthetic data (generated per Kuang et al. 2018) Amazon Review
Benchmarks: Synthetic Non-IID and IID OOD datasets Amazon Review cross-domain tasks (DEK→B, BEK→D, BDK→E, BDE→K)
Metrics: Test Accuracy RMSE F1 score
Baselines: Fed-FiS FPSO-FS EAMB-V3 EAMB-V5 CVS-V3 CVS-V5 PCFS-V3 PCFS-V5
Builds on: Fed-FiS FPSO-FS EAMB CVS PCFS
Closest papers by meaning
- Causality Inspired Federated Learning for OOD Generalization (ICML 2025 · similarity 0.88)
- A Simple Data Augmentation for Feature Distribution Skewed Federated Learning (CVPR 2025 · similarity 0.87)
- Causally Motivated Personalized Federated Invariant Learning with Shortcut-Averse Information-Theoretic Regularization (ICML 2024 · similarity 0.87)
- Federated Causal Structure Learning with Non-identical Variable Sets (ICML 2025 · similarity 0.87)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.86)
- FedCSL: A Scalable and Accurate Approach to Federated Causal Structure Learning (AAAI 2024 · similarity 0.86)
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.