Research map / Label-efficient and robust learning / Recognition and long-tail
Discovering Features with Synergistic Interactions in Multiple Views
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 SynFS, a deep embedded multi-view feature selector that uses interaction information and Bernoulli/Gaussian relaxed gates to find synergistic and non-synergistic feature subsets across views.
Problem
Selecting features whose interactions across multiple views are synergistic (more information jointly than separately) or non-synergistic, which existing multi-view and single-view feature selection methods do not capture.
Approach
Formalizes multi-view synergistic and non-synergistic feature selection via interaction information, relaxes the subset search with stochastic Gaussian-based Bernoulli gates, and trains view-specific synergistic and non-synergistic selectors with separate predictors, a mutual-exclusivity similarity loss, and a union-subset predictor.
Key results
- On synthetic data, SynFS reaches J-Index 1.00 with TPR 100 and FDR 0 across Syn1-Syn4, while CompFS drops to 0.40-0.48 on Syn2 and Syn3.
- On METABRIC, SynFS is the only method with positive SI(S) (0.36 ± 0.13) and reaches AUROC 0.862 using about 6% of features.
- On TCGA, SynFS achieves AUROC 0.683 versus 0.697 for MLP on all features, with 154 selected features.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SynFS instead of CompFS for interaction discovery | Feature selection · Syn1, Syn2, Syn3, and Syn4 | J-Index: Syn1: 0.91 ± 0.18 -> 1.00 ± 0.00; Syn2: 0.48 ± 0.25 -> 1.00 ± 0.00; Syn3: 0.40 ± 0.20 -> 1.00 ± 0.00; Syn4: 0.90 ± 0.20 | Helped |
| Use SynFS instead of traditional feature selection for interaction discovery | Feature selection · Synthetic datasets Syn1–Syn4 | J-Index: SynFS: 1.00 ± 0.00; STG: 0.47 ± 0.07†, 0.50 ± 0.00†, 0.50 ± 0.00†, 1.00 ± 0.00†; RForest: 0.50 ± 0.00†, 0.33 ± 0.00†, 0. | Helped |
| Use SynFS instead of CompFS for predictive performance | Feature selection · METABRIC | AUROC: 0.846 ± 0.03 -> 0.862 ± 0.02 | Helped |
| Use SynFS instead of STG or MLP for predictive performance | Feature selection · METABRIC | AUROC: STG: 0.887 ± 0.01; MLP: 0.884 ± 0.01; SynFS: 0.862 ± 0.02 | Hurt |
| Use SynFS instead of CompFS for predictive performance | Feature selection · TCGA | AUROC: CompFS: 0.686 ± 0.03; SynFS: 0.683 ± 0.03 | Mixed |
| Use SynFS instead of CompFS for synergistic interaction score | Feature selection · TCGA | SI(S): −2.27 ± 0.59 -> 7.87 ± 18.80 | Helped |
| Use SynFS instead of CompFS for predictive performance | Feature selection · PBMC | AUROC: 0.987 ± 0.00 -> 0.984 ± 0.00 | Hurt |
| Use SynFS instead of MLP for predictive performance | Feature selection · PBMC | AUROC: 0.977 ± 0.00 -> 0.984 ± 0.00 | Helped |
Surprising findings
- STG and RForest achieve perfect important-feature discovery on synthetic data but poor normalized J-Index because they cannot capture cross-view interactions.
Limitations
- Real-world interaction findings rely on literature support rather than verified ground truth.
- Unknown-ground-truth evaluation depends on a proposed Shapley-based SI metric whose validity is established mainly on synthetic data.
Experimental setup
Datasets: synthetic datasets (Chen et al., 2018; Imrie et al., 2022) MNIST METABRIC TCGA PBMC
Benchmarks: synthetic Syn1-Syn4 MNIST 3-vs-8 METABRIC TCGA PBMC
Metrics: Jaccard index normalized J-Index TPR FDR AUROC Set Interaction score (SI)
Baselines: CompFS STG Random Forest MLP
Builds on: CompFS STG (stochastic gates) Shapley value
Closest papers by meaning
- InfMasking: Unleashing Synergistic Information by Contrastive Multimodal Interactions (NeurIPS 2025 · similarity 0.84)
- Exploring the cloud of feature interaction scores in a Rashomon set (ICLR 2024 · similarity 0.84)
- On the Joint Interaction of Models, Data, and Features (ICLR 2024 · similarity 0.83)
- Estimating Conditional Mutual Information for Dynamic Feature Selection (ICLR 2024 · similarity 0.83)
- EASEMVC:Efficient Dual Selection Mechanism for Deep Multi-View Clustering (CVPR 2025 · similarity 0.82)
- Unsupervised Kernel-based Multi-view Feature Selection with Robust Self-representation and Binary Hashing (AAAI 2025 · similarity 0.82)
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.