Research map / Probabilistic and scientific modeling / Molecules and proteins
Towards Attributions of Input Variables in a Coalition
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.
Extends Shapley value via AND-OR interaction reallocation to define attributions for variable coalitions, explains attribution conflicts theoretically, and proposes three coalition faithfulness metrics validated on toy functions, NLP sentiment models, image classifiers, and KataGo.
Problem
Attribution methods such as Shapley values compute attributions for a predefined partition of input variables and offer no theory for forming meaningful partitions; attributions of a coalition often differ from the sum of its members' attributions.
Approach
Reformulates Shapley and Banzhaf values as allocations of AND-OR interaction effects, defines a coalition attribution by reallocating interactions that cover the coalition, and shows the conflict comes exactly from interactions that cover only part of the coalition.
Key results
- Coalition attribution reconstructs Shapley values with approximation errors around 1e-7 across coalition sizes 1-10.
- On toy functions, purely faithful coalitions score R(i)≈0.944 and Q(S)≈0.948, versus ≈0.031 and 0.013 for purely unfaithful ones.
- On SST-2, the coalition 'mesmerizing performances' gets Q=0.743 (BERT-large) while the unfaithful 'rivaling blair' gets Q=0.425.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed coalition faithfulness metrics to score known coalition types | MLP trained on toy functions · Toy functions, purely faithful versus purely unfaithful coalitions | R(i): 0.031 -> 0.944 | Helped |
| Use the proposed coalition faithfulness metrics to score known coalition types | MLP trained on toy functions · Toy functions, purely faithful versus purely unfaithful coalitions | R'(i): 0.016 -> 0.936 | Helped |
| Use the proposed coalition faithfulness metrics to score known coalition types | MLP trained on toy functions · Toy functions, purely faithful versus purely unfaithful coalitions | Q(S): 0.013 -> 0.948 | Helped |
| Select the natural phrase “mesmerizing performances” as a coalition rather than the phrase “rivaling blair” | Coalition faithfulness metrics on BERT-large · SST-2 sentence examples | Q(S): 0.425 -> 0.743 | Mixed |
| Use coalition attribution to approximate individual Shapley values | DNNs · DNN approximation of Shapley value, coalition sizes |S|=1–5 | Mean absolute approximation error: 3.6×10^{-8} to 7.6×10^{-7} | Helped |
| Use coalition attribution with partial-interaction effects to approximate model output | DNNs · DNN approximation of model output, coalition sizes |S|=1–5 | Mean absolute approximation error: 2.3×10^{-7} to 9.1×10^{-7} | Helped |
Surprising findings
- Coalitions that split a semantic unit (e.g., 'rivaling blair' splitting 'blair witch') are flagged as unfaithful, suggesting representation flaws in the model.
Limitations
- Interaction computation is computationally expensive, limiting experiments to small input sets (e.g., 10 stones in Go).
- Evaluations on NLP use only 10 manually selected words and a small set of annotated coalitions.
Experimental setup
Datasets: SST-2 MNIST CIFAR-10 synthetic toy functions
Benchmarks: SST-2 MNIST CIFAR-10 synthetic toy functions KataGo Go board positions
Metrics: R(i) R'(i) Q(S) approximation error of Shapley estimate approximation error of output mimicry
Baselines: Shapley value Banzhaf value Faith-Shap Joint Shapley value
Builds on: Shapley value Banzhaf value AND-OR interactions Faith-Shap
Closest papers by meaning
- On Blame Attribution for Accountable Multi-Agent Sequential Decision Making (NeurIPS 2021 · similarity 0.86)
- Deviate or Not: Learning Coalition Structures with Multiple-bit Observations in Games (AAAI 2025 · similarity 0.83)
- $\texttt{dattri}$: A Library for Efficient Data Attribution (NeurIPS 2024 · similarity 0.83)
- AttriBoT: A Bag of Tricks for Efficiently Approximating Leave-One-Out Context Attribution (ICLR 2025 · similarity 0.83)
- Explaining Probabilistic Models with Distributional Values (ICML 2024 · similarity 0.83)
- When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · 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.