marginal_effect: Get pseudo marginal effects from any model.

In DexGroves/hacktoolkit: Generic Hackathon Sauce

Description

Works by setting everyone in the data to a range of values in the variable of interest and scoring repeatedly. Very slow. Only compatible with data.tables.

Usage

marginal_effect(data, var, model, predict_model = predict, range = -1,
  sample_frac = 1)

Arguments

data	data.table object on which to score
var	string naming the variable of interest
model	target model object
predict_model	a function with signature (model, data) => predictions. Defaults to predict.
range	of values to test. Defaults to 10 evenly spaced points between the min and max of the target variable.
sample_frac	fraction of data to sample randomly

Value

data.table of the variable and its derived marginal effect

Examples

library("xgboost")
library("ggplot2")
library("foreach")
library("doMC")

data(diamonds)

set.seed(1234)

predict_model <- function(model, data) {
  score_mm <- model.matrix(price ~ ., data)
  score_dm <- xgb.DMatrix(score_mm)
  predict(model, newdata = score_dm)
}

train <- data.table(diamonds[order(runif(nrow(diamonds))), ][1:10000, ])
train_mm <- model.matrix(price ~ ., train)
train_y  <- train$price
train_dm <- xgb.DMatrix(train_mm, label = train$price)

model <- xgb.train(params = list(max_depth = 2,
                                 subsample = 0.5,
                                 eta = 0.1,
                                 colsample_bytree = 0.5,
                                 objective = "reg:linear"),
                   nrounds = 10,
                   data = train_dm)

marginal_effect(train, "x", model, predict_model)

DexGroves/hacktoolkit documentation built on May 6, 2019, 2:12 p.m.