SHAP

SHAP (SHapley Additive exPlanations) is a mainstream machine learning model explanation framework whose theoretical basis derives from Shapley values in game theory.

LIME and SHAP are methods in the field of explainable AI (XAI) that attempt to explain model behavior post hoc.

LIME (Local Interpretable Model-Agnostic Explanations) and SHAP (SHapley Additive exPlanations) are the primary tools for providing post-hoc explanations for complex black-box models.

SHAP基于博弈论中的Shapley值，能够量化每个特征对最终预测结果的贡献度，是目前工业界应用最广泛的解释性工具之一

LIME explains individual predictions by generating a perturbed dataset around the input sample to train a local linear approximation model, while SHAP provides feature attribution based on game-theoretic Shapley values; both are model-agnostic techniques.

Post-hoc explanation methods include SHAP and LIME, and intrinsic interpretable models include linear models and decision trees.

The study uses SHAP to quantify each feature's contribution to emotion prediction and GRAD-CAM for qualitative evaluation of the learned representations of vocal techniques, the two complementing each other to form a complete explanation chain from input features to model decisions.

Exact SHAP computation grows exponentially with the number of features, so in practice approximation algorithms such as KernelSHAP or TreeSHAP are typically used.

SHAP analysis shows that the effect of vibrato on emotion is context-dependent, specifically depending on articulation rate, so the same vibrato may convey different emotional coloring at different speech rates.

Current mainstream interpretability approaches include post-hoc attribution methods (such as attention visualization, SHAP) and mechanistic interpretability (such as sparse autoencoders, circuit analysis), with the latter attempting to locate identifiable "feature circuits" at the level of neuron activations.

Insufficient sampling times for SHAP values leads to unstable attribution results, manifested as fluctuations in feature rankings across runs; it is recommended to increase the sampling times or use DeepSHAP, and to report confidence intervals for the attributions.

SHAP values measure the contribution of features to model predictions and are based on statistical correlation rather than causal effects, so they cannot prove that temperature is a direct causal driver of strength.

SHAP and PDP analyses indicate that temperature is the primary factor affecting compressive strength after high-temperature exposure, followed in order by fine aggregate, water reducer, cement content, water-to-binder ratio, and water content.

Explainability techniques include feature importance analysis, local approximation explanations (such as LIME and SHAP), and inherently interpretable model design.

In 2017, Lundberg and Lee proposed SHAP, introducing the Shapley value into the field of machine learning interpretability.

Providing interpretability information (such as feature attributions generated by LIME or SHAP) can effectively counter automation bias.

Pixel-level attribution methods such as LIME, SHAP, and Grad-CAM use pixels or regions as the unit of explanation rather than semantic concepts familiar to humans.

SHAP and LIME are both post-hoc explanation frameworks that do not modify the model itself, but instead generate explanations after predictions are made.

Commonly used interpretability mitigation measures for medical AI in the industry include attention visualization (Grad-CAM), SHAP value explanations, and "human-in-the-loop" design.

The research integrated four feature importance methods — SHAP, grouped permutation importance, Boruta, and Recursive Feature Elimination (RFE) — and merged the results through a consensus ranking strategy.