Research map / Label-efficient and robust learning / Recognition and long-tail
Exploring the cloud of feature interaction scores in a Rashomon set
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.
Introduces feature interaction score (FIS) and its cloud (FISC) over a Rashomon set of near-equally accurate models, with a greedy sampling algorithm and Halo/swarm visualizations to show interaction strengths vary across accurate models.
Problem
Feature interactions extracted from a single trained predictive model may not be trustworthy, since multiple equally accurate models (a Rashomon set) can differ substantially in their feature interaction strengths.
Approach
Defines the feature interaction score as the loss change from jointly permuting features minus the sum of individual permutation loss changes, and characterizes the spread of these scores (FISC) across a mask-based Rashomon set sampled by a greedy search over input masks; visualizes it with Halo and swarm plots.
Key results
- On synthetic functions F2-F4, the FIS-in-context method achieves AUC 1.0 matching ArchDetect, while Two-way ANOVA, NID, SII and STI drop to roughly 0.5.
- On COMPAS, the greedy sampling yields a broader FISC range than AWP and random sampling for all listed pairs (e.g., age vs race FIS_max 0.002069 vs 0.000113 for AWP).
- In the MXene application, FIS shows a broad range when category Z is involved in interactions for induced charge prediction (reference R^2 0.98).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FIS in the specified interaction-detection context instead of Two-way ANOVA, Neural Interaction Detection, or Shapley interaction methods | Synthetic interaction detection · Synthetic F2–F4 | interaction detection AUC: F2: 0.51 -> 1.0; 0.52 -> 1.0; 0.50 -> 1.0; F3: 0.51 -> 1.0; 0.48 -> 1.0; 0.50 -> 1.0; F4: 0.55 -> 1.0; 0.56 -> 1.0; 0.51 | Helped |
| Use FIS in the specified interaction-detection context instead of ArchDetect | Synthetic interaction detection · Synthetic F1–F4 | interaction detection AUC: 1.0 vs 1.0 | No clear effect |
| Use greedy sampling rather than AWP or random sampling to explore the Rashomon set | FISC estimation with an MLP · Recidivism, age vs race | FIS range endpoints: AWP: -0.00026, 0.000113; greedy: -0.00089, 0.002069; random: -8.04E-06, 6.32E-06 | Helped |
Surprising findings
- Sampled models in the Rashomon set can have a lower FIS for one feature pair than another pair in a way that contradicts the ground-truth interaction calculated from the data-generating quadratic formula, although the FISC range covers the truth.
- The VIC compensation structure fails to capture the higher-order information compensation found by FISC (prior-charge interaction on the MLP/logistic setting).
Limitations
- Mask-based Rashomon set is one specific characterization; other ways of defining the Rashomon set are not explored.
- Visualizing interactions beyond three dimensions with Halo plots is not feasible without additional encoding or dimensionality reduction.
- Theoretical bounds of FISC are derived only for a specific sigmoid MLP under assumptions, with numerical root-finding; general case is approximated empirically.
Experimental setup
Datasets: COMPAS recidivism dataset (7,214 defendants, Broward County) synthetic functions F1-F4 quadratic equation synthetic data MXene intercalation dataset (360 compounds) ImageNet COCO
Benchmarks: synthetic interaction detection AUC (F1-F4) COMPAS recidivism prediction ImageNet image classification (ViT-B-16)
Metrics: interaction detection AUC model class reliance FIS min/max R^2 MSE
Baselines: Two-way ANOVA Neural Interaction Detection Shapley Interaction Index Shapley Taylor Interaction Index ArchDetect adversarial weight perturbation sampling random seed sampling
Builds on: Archipelago model reliance variable importance cloud Neural Interaction Detection
Closest papers by meaning
- On the Joint Interaction of Models, Data, and Features (ICLR 2024 · similarity 0.86)
- Dropout-Based Rashomon Set Exploration for Efficient Predictive Multiplicity Estimation (ICLR 2024 · similarity 0.85)
- Discovering Features with Synergistic Interactions in Multiple Views (ICML 2024 · similarity 0.84)
- Position: Amazing Things Come From Having Many Good Models (ICML 2024 · similarity 0.84)
- RashomonGB: Analyzing the Rashomon Effect and Mitigating Predictive Multiplicity in Gradient Boosting (NeurIPS 2024 · similarity 0.84)
- shapiq: Shapley Interactions for Machine Learning (NeurIPS 2024 · 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.