Pawlak Dominik HW4

Homeworks/Homework-III/Pawlak Dominik/Pawlak_HW3.Rmd

@@ -0,0 +1,101 @@
+---
+title: "WB_HW2"
+output: html_document
+date: '2022-05-17'
+---
+
+
+```{r setup, include=FALSE}
+knitr::opts_chunk$set(echo = TRUE)
+```
+
+
+
+```{r cars}
+library(DALEX)
+library(caret)
+set.seed(2137)
+
+data <- read.csv("insurance.csv")
+data$sex <- as.factor(data$sex)
+data$smoker <- as.factor(data$smoker)
+data$region <- as.factor(data$region)
+head(data)
+```
+Now, let's split the data into training and test datasets.
+```{r}
+set.seed(11)
+index <- createDataPartition(apartments$m2.price, p = 0.5, list = FALSE)
+
+train <- data[index,]
+test  <- data[-index,]
+
+x_train <- train[,-c(7)]
+y_train <- train[, 7]
+
+x_test <- test[,-c(7)]
+y_test <- test[, 7]
+```
+
+After splitting the data, we can train the models.
+```{r}
+library(ranger)
+library(gbm)
+
+forest <- ranger(charges~., data=train)
+y_pred <- predict(forest, x_test)
+print(y_pred$predictions[50])
+print(y_test[50])
+
+lr_model <- lm(charges~., data = train)
+pred_lr <- predict(lr_model, test)
+print(pred_lr[50])
+
+gbm_model <- gbm(charges~., data = train, distribution = "gaussian")
+y_gbm <- predict(gbm_model, x_test)
+print(y_gbm[50])
+```
+Lets's create explainers:
+```{r}
+explainer_rf <- DALEX::explain(forest, 
+                               data = x_test,  
+                               y = y_test,
+                               label = "random forest")
+
+explainer_lr <- DALEX::explain(lr_model, 
+                               data = x_test,  
+                               y = y_test,
+                               label = "linear regression")
+
+explainer_gbm <- DALEX::explain(gbm_model, 
+                               data = x_test,  
+                               y = y_test,
+                               label = "GBM model")
+```
+Let's create CP profiles for the 152th observation
+```{r}
+cp_pr <- predict_profile(explainer = explainer_rf,
+                         new_observation = x_test[152,])
+plot(cp_pr)
+plot(cp_pr, variables = c("sex", "smoker", "region"), variable_type = "categorical", 
+     categorical_type = "bars") 
+
+cp_lr <- predict_profile(explainer = explainer_lr,
+                         new_observation = x_test[152,])
+plot(cp_lr)
+plot(cp_lr, variables = c("sex", "smoker", "region"), variable_type = "categorical", 
+     categorical_type = "bars") 
+
+cp_gbm <- predict_profile(explainer = explainer_gbm,
+                         new_observation = x_test[152,])
+plot(cp_gbm)
+plot(cp_gbm, variables = c("sex", "smoker", "region"), variable_type = "categorical", 
+     categorical_type = "bars") 
+```
+All models suggest that the higher is age, the prediction is higher. However, the random forest curve has ups and dows and gbm curve looks a bit like steps and is almost never decreasing.
+
+As for BMI, the models aren't compatible either. The random forest curve first increases, than decreases and again increases. The gbm curve first increases and suddenly becomes constant.
+
+The number of children is almost linear in gbm model and doesn't seem to have big influence on the predictioon.
+
+All models claim that region and female variables don't have bigger influence. However the positive value in smoker variable strongly increases the prediction

Homeworks/Homework-IV/Pawlak Dominik/WB_HW4.Rmd

@@ -0,0 +1,84 @@
+---
+title: "WB_HW4"
+output: html_document
+date: '2022-05-25'
+---
+
+
+```{r setup, include=FALSE}
+knitr::opts_chunk$set(echo = TRUE)
+```
+
+
+
+```{r cars}
+library(DALEX)
+library(caret)
+set.seed(2137)
+
+data <- read.csv("insurance.csv")
+data$sex <- as.factor(data$sex)
+data$smoker <- as.factor(data$smoker)
+data$region <- as.factor(data$region)
+head(data)
+```
+Now, let's split the data into training and test datasets.
+```{r}
+set.seed(11)
+index <- createDataPartition(apartments$m2.price, p = 0.5, list = FALSE)
+
+train <- data[index,]
+test  <- data[-index,]
+
+x_train <- train[,-c(7)]
+y_train <- train[, 7]
+
+x_test <- test[,-c(7)]
+y_test <- test[, 7]
+```
+
+After splitting the data, we can train the models.
+```{r}
+library(ranger)
+library(gbm)
+
+forest <- ranger(charges~., data=train)
+y_pred <- predict(forest, x_test)
+print(y_pred$predictions[50])
+print(y_test[50])
+
+lr_model <- lm(charges~., data = train)
+pred_lr <- predict(lr_model, test)
+print(pred_lr[50])
+
+gbm_model <- gbm(charges~., data = train, distribution = "gaussian")
+y_gbm <- predict(gbm_model, x_test)
+print(y_gbm[50])
+```
+Lets's create explainers:
+```{r}
+explainer_rf <- DALEX::explain(forest, 
+                               data = x_test,  
+                               y = y_test,
+                               label = "random forest")
+
+explainer_lr <- DALEX::explain(lr_model, 
+                               data = x_test,  
+                               y = y_test,
+                               label = "linear regression")
+
+explainer_gbm <- DALEX::explain(gbm_model, 
+                               data = x_test,  
+                               y = y_test,
+                               label = "GBM model")
+```
+
+Once we have the explainers, let's calculate the variable importance measure for our models and plot them.
+```{r}
+vip_rf  <- model_parts(explainer_rf)
+vip_lr  <- model_parts(explainer_lr)
+vip_gbm <- model_parts(explainer_gbm)
+
+plot(vip_rf, vip_lr, vip_gbm)
+```
+All models suggest that 'smoker' variable is the most important. They all have similar value (around 14000-15000). GBM and Linear Regression models claim that age variable is second most important. However, random forest claims that bmi is more important than age. Two previous models place bmi variable on the third place. They all agree that remaining three variables are of no big importance. What is interesting, is that the initial value for GBM and linear regression is around 6200 and for random forest is around 5000.