waterfall: Explain prediction of a single observation

In EIX: Explain Interactions in 'XGBoost'

Description

This function calculates a table with influence of variables and interactions on the prediction of a given observation. It supports only xgboost models.

Usage

waterfall(
  xgb_model,
  new_observation,
  data,
  type = "binary",
  option = "interactions",
  baseline = 0
)

Arguments

xgb_model	a xgboost model.
new_observation	a new observation.
data	row from the original dataset with the new observation to explain (not one-hot-encoded). The param above has to be set to merge categorical features. If you dont wont to merge categorical features, set this parameter the same as new_observation.
type	the learning task of the model. Available tasks: "binary" for binary classification or "regression" for linear regression.
option	if "variables", the plot includes only single variables, if "interactions", then only interactions. Default "interaction".
baseline	a number or a character "Intercept" (for model intercept). The baseline for the plot, where the rectangles should start. Default 0.

Details

The function contains code or pieces of code from breakDown code created by Przemysław Biecek and xgboostExplainer code created by David Foster.

Value

an object of the broken class

Examples

library("EIX")
library("Matrix")
sm <- sparse.model.matrix(left ~ . - 1,  data = HR_data)

library("xgboost")
param <- list(objective = "binary:logistic", max_depth = 2)
xgb_model <- xgboost(sm, params = param, label = HR_data[, left] == 1, nrounds = 25, verbose=0)

data <- HR_data[9,-7]
new_observation <- sm[9,]

wf <- waterfall(xgb_model, new_observation, data,  option = "interactions")
wf

plot(wf)

EIX documentation built on March 23, 2021, 9:06 a.m.