knitr::opts_chunk$set(fig.height = 6, fig.width = 6, fig.align = "center")
Package EIX
is the set of tools to explore the structure of XGBoost and lightGBM models. It includes functions finding strong interactions and also checking importance of single variables and interactions by usage different measures. EIX
consists several functions to visualize results.
Almost all EIX
functions require only two parameters: a XGBoost or LightGBM model and data table used as training dataset. The exceptions are the waterfall
function and its plot. The first one requires parameters: a XGBoost model and observation, which prediction has to be explained). These two functions support only XGBoost models. All plots are created with package ggplot2
. Most of them use plot theme theme_mi2
from DALEX
.
This vignette shows usage of EIX
package. It lets to explain XGBoost prediction model concerning departures of employees from company using HR_data. Dataset was taken from kaggle and consists 14999 observations and 10 variables. The dataset is also available in package EIX
and there it is described more precisely.
#devtools :: install_github("ModelOriented/EIX") library("EIX") set.seed(4) knitr::kable(head(HR_data))
To create correct XGBoost model, remember to change categorical features to factors and next change the data frame to sparse matrix. The categorical features are one-hot encoded.
library("Matrix") sparse_matrix <- sparse.model.matrix(left ~ . - 1, data = HR_data) head(sparse_matrix)
Package EIX
uses table, which was generated by xgboost::xgb.model.dt.tree
with information about trees, their nodes and leaves.
library("xgboost") param <- list(objective = "binary:logistic", max_depth = 2) xgb_model <- xgboost(sparse_matrix, params = param, label = HR_data[, left] == 1, nrounds = 50, verbose = FALSE) knitr::kable(head(xgboost::xgb.model.dt.tree(colnames(sparse_matrix),xgb_model)))
Function xgboost::xgb.importance
shows importance of single variables. EIX
adds new measures of variables’ importance and shows also importance of interactions.
knitr::kable(head(xgboost::xgb.importance(colnames(sparse_matrix),xgb_model)))
The lollipop
plot is used to visualize the model in such way that the most important variables and interactions are visible.
On the x-axis, there are tree numbers and on the y-axis there is Gain measure for each node. One segment is one tree in the model and each point is one node. On the plot there are all nodes, which are not leaves. Shape of points signifies depth of node. All roots on the plot are connected by a red line. If in the same segment there is a variable with a higher depth above the variable with a lower depth, it means that interaction occurs.
There is opportunity to set a different way of labeling. On the plot we can see the most important variables in roots (horizontal labels), and interactions (vertical labels), this is option labels = "topAll"
which is default. Moreover, there are two additional options: labels = "roots"
- for variables in roots only, labels = "interactions"
for interactions only. The numbers of labels visible on the plot you can change by parametr threshold
(range from 0 to 1, default 0.1).
The plot is on a logarithmic scale because the initial trees usually are the most important. You can change the scale of the plot by setting the parameter log_scale = FALSE
.
lolli<-lollipop(xgb_model,sparse_matrix) plot(lolli) #plot(lolli, threshold=0.05) #plot(lolli, labels="roots") #plot(lolli, labels="interactions") #plot(lolli, labels="roots", threshold=0.05) #plot(lolli, labels="interactions",threshold=0.05) #plot(lolli, log_scale = FALSE)
We can consider interactions in two ways. In first approach we can explore all pairs of variable, which occur in the model one above the other. This approach is not the best one, because we cannot distinguish if pair of variables are real interaction or not. In this approach high gain of pair can be a result of high gain of down variable (child). To explore pairs of variables you can generate table with them using function interactions
with parametr option = "pairs"
. This table includes Gain measure and number of occurrences of pairs. You can also use the function plot
to visualize Gain measure.
pairs<-interactions(xgb_model, sparse_matrix, option = "pairs") head(pairs) plot(pairs)
The interactions
plot is a matrix plot with a child from the pair on the x-axis and the parent on the y-axis. The color of the square at the intersection of two variables means value of sumGain measure. The darker square, the higher sumGain of variable pairs. The range of sumGain measure is divided into four equal parts: very low, low, medium, high
.
In second approach, to find strong interactions, we can consider only these pairs of variables, where variable on the bottom (child) has higher gain than variable on the top (parent). We can also create ranking of interactions using function importance
with parameter option = "interactions"
. More details in the next section.
interactions<-interactions(xgb_model, sparse_matrix, option = "interactions") head(interactions) plot(interactions)
For exploring variables’ and interactions’ importance there are three functions in EIX
package: importance
, its plot
with parameter radar = TRUE
or radar = FALSE
.
With EIX
package we can compare importance of single variables and interactions. The functions importance
can return three kinds of outputs, depending on the opt
parameter:
option = "variables"
- it consists only single variables
option = "interactions"
- only interactions
option = "both"
- output shows importance both single variables and interactions.
NOTE: option = "both"
is not direct connection option = "variables"
and option = "interactions"
, because values of variable importance measure, which were in the interactions, are not included in importance of single variable.
In EIX
the following measures are available:
EIX
package gives additionally measures of variables importance for single variable:
The function importance
returns a table with all available importance measures for given option.
The table is sorted by descending value of sumGain.
The function plot
with parameter radar = FALSE
and a result from the importance
function as an argument shows two measures of importance, which can be chosen by xmeasure
and ymeasure
parameters. By parameter top
we can decide how many positions will be included in the plot.
importance<-importance(xgb_model, sparse_matrix, option = "both") head(importance) plot(importance, radar=FALSE) #plot(importance, xmeasure = "mean5Gain", ymeasure = "sumGain", top = 15, radar=FALSE)
The function plot
with parameter radar = TRUE
enables to compare different measures of variables and interactions importance on the radar plot from ggiraphExtra
package.
Bellow I attach the example of radar plot. On the outside of the circle there are names of variables or interactions. Colored lines represent various measures of importance. The positions on the plot are sorted decreasing. The variable with the highest sumGain value is on the right of 12 o'clock. Next the sumGain value decreases in a clockwise direction. On the plot it is possible to change place, where the features names start by parameter text_start_point
(range from 0 to 1, default 0.5), and size of this text by parametrer text_size
.
plot(importance) #plot(importance, text_start_point = 0.3) #plot(importance, text_size = 4) #plot(importance, top=15)
For single prediction explaining package EIX
uses two packages: xgboostExplainer
i breakDown
. The package xgboostExplainer
is a tool to interpreting prediction of xgboost model. The package EIX
uses its code and modifies it to include interactions. The methodology of plot creation comes from package breakDown
.
The function waterfall
returns table with variables’ impact on the prediction of the model. Depending on the parameter option
, the table includes interactions (option = "interactions"
- default) or does not (option = "variables"
). The function plot
with waterfall
object as an argument visualizes this table. On the y-axis there are: intercept (it is the probability that random variable from training dataset will be 1), variables (which have an impact on prediction) and final prognosis of the model. On the x-axis there is log-odds of impact each variables.
data <- HR_data[9,] new_observation <- sparse_matrix[9,] wf<-waterfall(xgb_model, new_observation, data, option = "interactions") wf plot(wf) #wf<-waterfall(xgb_model, new_observation, data, option = "interactions", baseline = "intercept") #wf #plot(wf)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.