Probabilistic Linear Explanations
The paper proposes sparse, anchored linear explanations that work for both classification and regression.
The framework maps instances onto a Boolean hypercube so explanations can show the size and direction of feature contributions under a fixed sparsity budget. The authors prove relevance-error minimization is hard for neural-network models, then connect it to a more tractable fidelity-error surrogate. They offer both an optimal MIP formulation and a faster Iterative Hard Thresholding algorithm with approximation guarantees. In experiments, their method reports lower relevance error than LIME and MAPLE while satisfying anchoring and sparsity constraints by construction. ArXiv · AI/CL/LG's note
The framework maps instances onto a Boolean hypercube so explanations can show the size and direction of feature contributions under a fixed sparsity budget. The authors prove relevance-error minimization is hard for neural-network models, then connect it to a more tractable fidelity-error surrogate. They offer both an optimal MIP formulation and a faster Iterative Hard Thresholding algorithm with approximation guarantees. In experiments, their method reports lower relevance error than LIME and MAPLE while satisfying anchoring and sparsity constraints by construction. ArXiv · AI/CL/LG's note
score 4