1.4. Generalized Total Sobol Index#
The Generalized Total Sobol Index (TSI) is a theoretical measure of feature importance that comes from sensitivity analysis. It has gained popularity in the field of interpretable Machine Learning (Bénard et al.[1], Williamson et al.[2]) as an interesting target to assess the relevance of features in predictive models. Indeed, the TSI can be intuitively defined by the performance drop of a predictive model when a feature is removed, similarly to an ablation study. More formally, it can be defined as:
where \(X^{-j}\) denotes the feature vector without the \(j^{th}\) feature, and \(\mu_{-j}(X^{-j})\) is the same predictive model as \(\mu(X)\) but retrained on the reduced feature set \(X^{-j}\). When \(\mathcal{L}\) is the squared loss, for a regression task, \(\mu_{-j}(X^{-j}) = \mathbb{E}[y \mid X^{-j}]\) and when \(\mathcal{L}\) is the log-loss, for a classification task, \(\mu_{-j}(X^{-j}) = P(y \mid X^{-j})\).
It is defined as generalized TSI since the original Sobol indices were introduced in the context of variance-based sensitivity analysis (Sobol[3]) which corresponds to having a quadratic loss.
In general, it can be estimated directly using a plug-in estimator in the definition, which gives the Leave-One-Covariate-Out (Leave-One-Covariate-Out).
1.4.1. Mean Squared Error (MSE) case#
In the regression setting with the MSE loss, the TSI can be expressed in several equivalent forms which provide different intuitions of why the target quantity is relevant:
The first formulation (loss / refitting) provides the definition of the TSI with the quadratic loss. The second formulation (adjustment) shows that the TSI can be seen as the dropped R-squared when removing the \(j^{th}\) feature, scaled by the output variance. The third formulation (variance) shows that it quantifies the variance of the output explained uniquely by the \(j^{th}\) feature. The fourth formulation (marginalization) shows that it can be seen as the performance drop when marginalizing the contribution of the \(j^{th}\) feature in the predictive model. Finally, the last formulation (perturbation) shows that it can be seen as the dropped performance by perturbing the \(j^{th}\) feature conditionally on the rest.
Also, we observe that LOCO (Leave-One-Covariate-Out), CFI (Conditional Feature Importance) and conditional SAGE value functions (Covert et al.[4]) target the TSI since they consists on plug-in estimators of the refitting, perturbation and marginalization formulation respectively.
1.4.2. Cross-entropy case#
In the classification setting with the cross-entropy (log-loss), the TSI can be expressed in terms of information-theoretic quantities:
From the first formulation (mutual information), we see that the TSI quantifies the mutual information between the output \(Y\) and the feature \(X^j\) conditionally on the rest of the features \(X^{-j}\). Thus, it quantifies how much information about \(Y\) is contained in \(X^j\) that is not already contained in \(X^{-j}\). The second formulation (Kullback-Leibler divergence) shows that the TSI can be seen as how much the original probability distribution diverges from the one that we would have obtained if the \(j^{th}\) feature was conditionally independent from the output given the rest of the features.