Q-LIME $π$: A Quantum-Inspired Extension to LIME
Nelson Colón Vargas
TL;DR
The paper tackles the computational burden of local explainability in high-dimensional spaces by introducing Q-LIME π, a quantum-inspired extension of LIME that encodes active binary features into a quantum-like state and uses bit-flip perturbations to probe feature contributions. By mapping perturbations to simple quantum-like operations, it aims to build a local surrogate explanation more efficiently than classical LIME. Experimental evidence on IMDb sentiment data shows that Q-LIME π often reproduces the top-5 feature rankings of classical LIME while achieving substantial runtime reductions in small-to-moderate feature spaces. This hybrid quantum-classical approach demonstrates a viable path toward scalable, interpretable AI and motivates further exploration as quantum hardware and methods advance.
Abstract
Machine learning models offer powerful predictive capabilities but often lack transparency. Local Interpretable Model-agnostic Explanations (LIME) addresses this by perturbing features and measuring their impact on a model's output. In text-based tasks, LIME typically removes present words (bits set to 1) to identify high-impact tokens. We propose \textbf{Q-LIME $π$} (Quantum LIME $π$), a quantum-inspired extension of LIME that encodes a binary feature vector in a quantum state, leveraging superposition and interference to explore local neighborhoods more efficiently. Our method focuses on flipping bits from $1 \rightarrow 0$ to emulate LIME's ``removal'' strategy, and can be extended to $0 \rightarrow 1$ where adding features is relevant. Experiments on subsets of the IMDb dataset demonstrate that Q-LIME $π$ often achieves near-identical top-feature rankings compared to classical LIME while exhibiting lower runtime in small- to moderate-dimensional feature spaces. This quantum-classical hybrid approach thus provides a new pathway for interpretable AI, suggesting that, with further improvements in quantum hardware and methods, quantum parallelism may facilitate more efficient local explanations for high-dimensional data.