In leylabmpi/endoR: Interpret and Visualize Stable Tree Ensemble Models

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width = 7,
  fig.height = 2.5
)

This vignette illustrates how to interpret multiclass models with endoR. We will use the iris data for this purpose: the species is being predicted (= target) using the length and width of petals and sepals (= features).

Preambule

library(endoR)
library(randomForest)
library(tidyverse)
library(caret)
library(ggpubr)
library(ggraph)

Data

data(iris)
summary(iris)

Fit model

set.seed(1313)
mod <- randomForest(Species ~ ., data = iris)
mod

Fit endoR for the "setosa" species

We will only prune (= TRUE) and discretize decisions in K = 3 levels.

endo_setosa <- model2DE(model = mod, model_type = 'rf'
                 , data = select(iris, -Species), target = iris$Species
                 , classPos = 'setosa' # our focal class
                 , filter = FALSE 

                 # we filter in K = 3 categories the numerical features
                 , discretize = TRUE, K = 3)

Plants from the setosa species have small petals and narrow, long sepals.

plotFeatures(endo_setosa, levels_order = c('Low', 'Medium', 'High'))

# The warnings du to the font are due to Windows .. no worries.
plotNetwork(endo_setosa, hide_isolated_nodes = FALSE, layout = 'fr')

The only interaction used for predictions is the one of sepal length (High = long) and width (Low = narrow). We can also use ggplot2 theme to format the plot, e.g., put the legend boxes next to each other.

plotNetwork(endo_setosa, hide_isolated_nodes = TRUE, layout = 'fr')+
  theme(legend.box = "horizontal")

Fit endoR for the "versicolor" species

This time we will filter decisions based on their importance to trim the network (filter = TRUE). We will use min_imp = 0.5 to keep at least all decisions with an importance > 0.5*the best importance (the lower min_imp, the slighter the filtering).

endo_versicolor <- model2DE(model = mod, model_type = 'rf'
                 , data = select(iris, -Species), target = iris$Species
                 , classPos = 'versicolor'
                 , K = 3, discretize = TRUE
                 , filter = TRUE, min_imp = 0.5)

The petal's proportions are intermediary compared to the setosa and virginica species.

plotFeatures(endo_versicolor, levels_order = c('Low', 'Medium', 'High'))

Sepals have a narrow-intermediary width and long-intermediary length (= if wide and small then it's not a versicolor, as seen on the network).

plotNetwork(endo_versicolor, hide_isolated_nodes = FALSE, layout = 'fr'
            # we show only edges that connect 3 nodes max -> removes edges with 
            # lowest importances - for longer paths = more complex network,
            # you can increase path_length
            , path_length = 3  
            )+
  scale_edge_alpha(range = c(0.8,1))+
  theme(legend.box = "horizontal")