EIX: Titanic data

Ewelina Karbowiak

2018-29-03

Data Info

This vignette shows usage of EIX package for titanic data. This dataset was copied from stablelearner package. With EIX package we explain XGBoost classification model concerning the survival problem. More details about EIX package here.

#devtools :: install_github("ModelOriented/EIX")
library("EIX")
library(data.table)
## Warning: package 'data.table' was built under R version 4.0.4
set.seed(4)
titanic_data<-data.table(na.omit(titanic_data))
knitr::kable(head(titanic_data))
gender age class embarked country fare sibsp parch survived
male 42 3rd Southampton United States 7.11 0 0 no
male 13 3rd Southampton United States 20.05 0 2 no
male 16 3rd Southampton United States 20.05 1 1 no
female 39 3rd Southampton England 20.05 1 1 yes
female 16 3rd Southampton Norway 7.13 0 0 yes
male 25 3rd Southampton United States 7.13 0 0 yes
library("Matrix")
sparse_matrix <- sparse.model.matrix(survived ~ . - 1,  data = titanic_data)

Xgboost model creation

library("xgboost")
param <- list(objective = "binary:logistic", max_depth = 2)
xgb_model <- xgboost(sparse_matrix, params = param, label = titanic_data[, "survived"] == "yes", nrounds = 50, verbose = FALSE)
## [00:13:12] WARNING: amalgamation/../src/learner.cc:1061: Starting in XGBoost 1.3.0, the default evaluation metric used with the objective 'binary:logistic' was changed from 'error' to 'logloss'. Explicitly set eval_metric if you'd like to restore the old behavior.

Model visualization

First let’s plot the model.

lolli<-lollipop(xgb_model,sparse_matrix)
plot(lolli, threshold=0.02)
## Warning: ggrepel: 5 unlabeled data points (too many overlaps). Consider
## increasing max.overlaps

Interactions

Next we explore interactions using interactions() functions and its plot.

interactions<-interactions(xgb_model, sparse_matrix, option = "interactions")
head(interactions, 15)
##                    Parent                 Child   sumGain frequency
##  1:              class3rd                  fare 45.577697         5
##  2:                 sibsp                   age 23.668270         3
##  3:          genderfemale                   age 23.227908         4
##  4:              class3rd                   age 22.997468         5
##  5:          genderfemale                  fare 13.975253         2
##  6:    countrySwitzerland                  fare 11.700150         2
##  7:                  fare classrestaurant staff  7.023322         1
##  8:                   age  countryUnited States  4.857722         1
##  9:    countrySwitzerland                   age  4.250423         2
## 10: classrestaurant staff              class3rd  4.190052         1
## 11:        classdeck crew                   age  4.165662         2
## 12:                  fare          genderfemale  4.100826         1
## 13:                   age                 sibsp  4.055897         1
## 14:                   age                  fare  4.045337         1
## 15:          genderfemale              class2nd  3.013065         1
plot(interactions)

Variables’ and interactions’ importance

importance<-importance(xgb_model, sparse_matrix, option = "both")
head(importance, 15)
##                     Feature sumGain sumCover meanGain meanCover frequency
##  1:            genderfemale  840.50 5042.000   60.030   360.100        14
##  2:                class3rd  205.10 2172.000   20.510   217.200        10
##  3:          classdeck crew  115.50 2102.000   19.250   350.300         6
##  4:                     age   90.00 3870.000    5.294   227.600        17
##  5:                    fare   58.59 2996.000    3.255   166.400        18
##  6:       embarkedCherbourg   48.64  787.500   24.320   393.800         2
##  7:           class3rd:fare   45.58  682.100    9.116   136.400         5
##  8:    countryUnited States   27.46  519.500   13.730   259.800         2
##  9:   classrestaurant staff   27.41 1970.000    4.569   328.300         6
## 10:               sibsp:age   23.67  684.900    7.889   228.300         3
## 11:        genderfemale:age   23.23  849.400    5.807   212.400         4
## 12:            class3rd:age   23.00  971.200    4.599   194.200         5
## 13:                   sibsp   19.56  921.200    6.519   307.100         3
## 14:       genderfemale:fare   13.98  348.900    6.988   174.500         2
## 15: countrySwitzerland:fare   11.70    6.044    5.850     3.022         2
##     mean5Gain
##  1:   159.700
##  2:    37.050
##  3:    22.900
##  4:    12.700
##  5:     6.433
##  6:    24.320
##  7:     9.116
##  8:    13.730
##  9:     5.247
## 10:     7.889
## 11:     5.807
## 12:     4.599
## 13:     6.519
## 14:     6.988
## 15:     5.850
plot(importance, radar=FALSE)

plot(importance)

Explanation of the single prediction including interactions

Let’s see an explanation of the prediction for an 18-year-old from England who has traveled 3rd class.

data <- titanic_data[27,]
new_observation <- sparse_matrix[27,]
wf<-waterfall(xgb_model, new_observation, data, option = "interactions")
wf
##                                 contribution
## xgboost: intercept                    -0.819
## xgboost: gender = 2                    1.714
## xgboost: class = 3                    -1.437
## xgboost: country:fare = 15:8.06       -1.011
## xgboost: country = 15                 -0.635
## xgboost: embarked = 4                  0.310
## xgboost: age:class = 18:3              0.119
## xgboost: class:age = 3:18             -0.116
## xgboost: fare = 8.06                  -0.103
## xgboost: age:country = 18:15           0.102
## xgboost: class:fare = 3:8.06           0.096
## xgboost: age = 18                     -0.094
## xgboost: age:embarked = 18:4           0.089
## xgboost: gender:class = 2:3           -0.043
## xgboost: sibsp:age = 0:18             -0.022
## xgboost: sibsp = 0                     0.021
## xgboost: gender:age = 2:18             0.020
## xgboost: fare:age = 8.06:18            0.019
## xgboost: gender:fare = 2:8.06         -0.018
## xgboost: age:sibsp = 18:0              0.010
## xgboost: parch:fare = 0:8.06           0.002
## xgboost: parch = 0                     0.001
## xgboost: prediction                   -1.794
plot(wf)