Research map / Probabilistic and scientific modeling / Molecules and proteins
CounterFlowNet: From Minimal Changes to Meaningful Counterfactual Explanations
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 CounterFlowNet, a conditional GFlowNet that samples diverse, sparse counterfactual explanations for tabular data with probability proportional to a composite validity/proximity/plausibility/sparsity reward, with inference-time action masking for actionability constraints.
Problem
Generating multiple high-quality counterfactual explanations for tabular data that are sparse, handle heterogeneous (continuous and categorical) features, and respect user-defined actionability constraints.
Approach
Frames counterfactual generation as sequential feature modification in a two-stage (feature index, then value) action space, trained with trajectory balance so that sampling probability is proportional to a non-differentiable composite CF-quality reward; constraints are enforced via action masking at inference time.
Key results
- Protocol A (German Credit): CFN reaches 51.65% harmonic mean vs 46.39% for the next best (L2C), with 99.48% validity and 100% coverage.
- Protocol B (Adult): CFN gets best proximity (0.19 vs 0.44 for DICE) and epsilon-sparsity (0.36 vs 0.50 for DICE), with diversity 0.21 versus DiCoFlex's 0.40.
- Inference-time action masking under constraint Set 1 on Adult matches retraining (validity 0.99, proximity 0.24 for both) while guaranteeing constraint satisfaction by construction.
Surprising findings
- On Protocol B, DiCoFlex achieves higher diversity than CounterFlowNet, and DICE/CCHVAE produce sparser categorical changes on several datasets.
Limitations
- Continuous features must be quantized into bins (B=4 or B=64), and reconstruction uses bin centers.
- Protocol B results use a single seed, and Protocol A results are averaged over five seeds.
- Diversity is weaker than DiCoFlex on Protocol B, and some baselines are stronger on categorical sparsity.
Experimental setup
Datasets: German Credit Adult Income Graduate Admission Student Performance Lending Club GMC Bank Marketing Default
Benchmarks: L2C protocol (Protocol A) Protocol B continuous/categorical counterfactual benchmark
Metrics: validity sparsity diversity harmonic mean coverage unary constraint satisfaction proximity epsilon-sparsity LOF
Baselines: DiCE COPA MCCE L2C C-CHVAE DiCoFlex
Builds on: GFlowNet trajectory balance DiCE L2C DiCoFlex
Closest papers by meaning
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.91)
- DiCoFlex: Model-Agnostic Diverse Counterfactuals with Flexible Control (NeurIPS 2025 · similarity 0.89)
- V-CECE: Visual Counterfactual Explanations via Conceptual Edits (NeurIPS 2025 · similarity 0.88)
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 2024 · similarity 0.87)
- Robust Stochastic Graph Generator for Counterfactual Explanations (AAAI 2024 · similarity 0.87)
- Counterfactual Maps: What They Are and How to Find Them (NeurIPS 2026 · similarity 0.87)
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.