Tous les livres

Professionnel

Applis À propos Coach Connexion Commencer la lecture

Quantitative Finance · Glossaire

Qu'est-ce que « Partial dependence, individual conditional expectation, accumulated local effects » ?

Aussi appelé : partial dependence · individual conditional expectation · accumulated local effects

Definition 21.2 Machine Learning for Markets · Chapitre 21 — Interpretability and Model Governance

The partial dependence of a model ff on feature jj is PDj(v)=n−1∑if(v,xi,−j)\mathrm{PD}_j(v) = n^{-1}\sum_i f(v, x_{i,-j}): the average prediction with feature jj set to vv for every observation (Friedman, 2001). The individual conditional expectation (ICE) curves are the terms of that average, one curve per observation (Goldstein and co-authors, 2015). Accumulated local effects (ALE) average, within narrow intervals of feature jj, the change in prediction when jj moves across the interval for the observations that lie in it, and accumulate those changes (Apley and Zhu, 2020).

correlation of x1x_1 and x2x_2
root mean square error against the true effect of x1x_100.920.99
partial dependence0.0270.1340.229
accumulated local effects0.0280.0440.110
Table 21.1. Error of two global explanations of the same fitted model, evaluated at the data’s values of x1x_1 (the true effect ranges over 2). Data: ml_explain.effects.
The effect of x_1 on a gradient-boosted model when x_2 is a copy of x_1 with correlation 0.99. Data: ml_explain.curves.
Figure 21.1. The effect of x1x_1 on a gradient-boosted model when x2x_2 is a copy of x1x_1 with correlation 0.99. Data: ml_explain.curves.
Lire dans le chapitre →