Customer Segment Prediction Using Machine Learning - Logistic Regression and KNN Methods
Claudy Sonly Daniel
Updated: December 26 2018
First, let’s set up cathing for this notebook and load pacman package for loading libraries conveniently.
knitr::opts_chunk$set(cache = TRUE)
options(scipen = 9999)
if (!require("pacman")) install.packages("pacman")## Loading required package: pacmanThen, load the data.
p_load(dplyr)
wholesale <- read.csv("data_input/wholesale.csv") %>%
select(-2) # remove region feature
str(wholesale)## 'data.frame': 440 obs. of 7 variables:
## $ Channel : int 2 2 2 1 2 2 2 2 1 2 ...
## $ Fresh : int 12669 7057 6353 13265 22615 9413 12126 7579 5963 6006 ...
## $ Milk : int 9656 9810 8808 1196 5410 8259 3199 4956 3648 11093 ...
## $ Grocery : int 7561 9568 7684 4221 7198 5126 6975 9426 6192 18881 ...
## $ Frozen : int 214 1762 2405 6404 3915 666 480 1669 425 1159 ...
## $ Detergents_Paper: int 2674 3293 3516 507 1777 1795 3140 3321 1716 7425 ...
## $ Delicassen : int 1338 1776 7844 1788 5185 1451 545 2566 750 2098 ...Here, first we proceed with KNN method. Let’s convert Channel feature into Industry, and make it factor.
wholesale.k <- wholesale
wholesale.k <- wholesale %>%
mutate(Industry = factor(Channel, levels = c(1, 2), labels = c("horeca", "retail"))) %>%
select(-1) # remove the original Channel feature
wholesale.k %>% head()## Fresh Milk Grocery Frozen Detergents_Paper Delicassen Industry
## 1 12669 9656 7561 214 2674 1338 retail
## 2 7057 9810 9568 1762 3293 1776 retail
## 3 6353 8808 7684 2405 3516 7844 retail
## 4 13265 1196 4221 6404 507 1788 horeca
## 5 22615 5410 7198 3915 1777 5185 retail
## 6 9413 8259 5126 666 1795 1451 retailCheck the proportion of the Industry feature.
round(prop.table(table(wholesale.k$Industry)) * 100, 2)##
## horeca retail
## 67.73 32.27Normalize the data using z-score normalization.
wholesale.z <- as.data.frame(scale(wholesale.k[, -7]))
summary(wholesale.z)## Fresh Milk Grocery Frozen
## Min. :-0.9486 Min. :-0.7779 Min. :-0.8364 Min. :-0.62763
## 1st Qu.:-0.7015 1st Qu.:-0.5776 1st Qu.:-0.6101 1st Qu.:-0.47988
## Median :-0.2764 Median :-0.2939 Median :-0.3363 Median :-0.31844
## Mean : 0.0000 Mean : 0.0000 Mean : 0.0000 Mean : 0.00000
## 3rd Qu.: 0.3901 3rd Qu.: 0.1889 3rd Qu.: 0.2846 3rd Qu.: 0.09935
## Max. : 7.9187 Max. : 9.1732 Max. : 8.9264 Max. :11.90545
## Detergents_Paper Delicassen
## Min. :-0.6037 Min. :-0.5396
## 1st Qu.:-0.5505 1st Qu.:-0.3960
## Median :-0.4331 Median :-0.1984
## Mean : 0.0000 Mean : 0.0000
## 3rd Qu.: 0.2182 3rd Qu.: 0.1047
## Max. : 7.9586 Max. :16.4597Combine the columns again.
wholesale.n <- as.data.frame(cbind(wholesale.z, "Industry" = wholesale.k$Industry))
summary(wholesale.n)## Fresh Milk Grocery Frozen
## Min. :-0.9486 Min. :-0.7779 Min. :-0.8364 Min. :-0.62763
## 1st Qu.:-0.7015 1st Qu.:-0.5776 1st Qu.:-0.6101 1st Qu.:-0.47988
## Median :-0.2764 Median :-0.2939 Median :-0.3363 Median :-0.31844
## Mean : 0.0000 Mean : 0.0000 Mean : 0.0000 Mean : 0.00000
## 3rd Qu.: 0.3901 3rd Qu.: 0.1889 3rd Qu.: 0.2846 3rd Qu.: 0.09935
## Max. : 7.9187 Max. : 9.1732 Max. : 8.9264 Max. :11.90545
## Detergents_Paper Delicassen Industry
## Min. :-0.6037 Min. :-0.5396 horeca:298
## 1st Qu.:-0.5505 1st Qu.:-0.3960 retail:142
## Median :-0.4331 Median :-0.1984
## Mean : 0.0000 Mean : 0.0000
## 3rd Qu.: 0.2182 3rd Qu.: 0.1047
## Max. : 7.9586 Max. :16.4597Divide to train and test set.
set.seed(2)
index_sample <- sample(nrow(wholesale.n))
wholesale_train <- wholesale.n[index_sample, ]
wholesale_test <- wholesale.n[index_sample, ]
prop.table(table(wholesale_train$Industry))##
## horeca retail
## 0.6772727 0.3227273prop.table(table(wholesale_test$Industry))##
## horeca retail
## 0.6772727 0.3227273Create accuracy function.
accuracy <- function(data) {
# initialize number of predictions
correct <- 0
for (i in c(1:nrow(data))) {
# 7th variable is actual class, 8th is our predicted class
if (data[i, 7] == data[i, 8]) {
correct <- correct + 1
}
}
percentage <- correct / nrow(data) * 100
return(percentage)
}Finally create knn prediction, and check the accuracy.
p_load(class)
predictions <- knn(train = wholesale_train[, 1:6], test = wholesale_test[, 1:6], cl = wholesale_train[, 7], k = 19)
wholesale_test[, 8] <- predictions
print(accuracy(wholesale_test))## [1] 90.90909Let’s try another method to predict, which is logistic regression. We have to assure there is no multicollinearity between all variables. Here I use chart.Correlation function to see the correlation between variables.
p_load(PerformanceAnalytics)
wholesale.n[, 7] <- as.numeric(wholesale.n[, 7])
chart.Correlation(wholesale.n)
It looks like Grocery has great correlation with Detergents_Paper and less affecting to Industrythan Detergents_Paper. So, let’s remove it, and proceed step-wise regression for feature selection with backward elimination method to find the best predictors for the model.
lm.all <- glm(Industry ~ Fresh + Milk + Frozen + Detergents_Paper + Delicassen, data = wholesale_train, family = "binomial")## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurredstep(lm.all, direction = "backward")## Start: AIC=220.15
## Industry ~ Fresh + Milk + Frozen + Detergents_Paper + Delicassen## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred## Df Deviance AIC
## - Delicassen 1 208.17 218.17
## - Fresh 1 208.40 218.40
## <none> 208.15 220.15
## - Frozen 1 214.11 224.11
## - Milk 1 214.48 224.48
## - Detergents_Paper 1 348.67 358.67## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred##
## Step: AIC=218.17
## Industry ~ Fresh + Milk + Frozen + Detergents_Paper## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred## Df Deviance AIC
## - Fresh 1 208.40 216.40
## <none> 208.17 218.17
## - Milk 1 214.69 222.69
## - Frozen 1 215.54 223.54
## - Detergents_Paper 1 348.74 356.74## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred##
## Step: AIC=216.4
## Industry ~ Milk + Frozen + Detergents_Paper## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred
## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurred## Df Deviance AIC
## <none> 208.40 216.40
## - Milk 1 215.38 221.38
## - Frozen 1 215.76 221.76
## - Detergents_Paper 1 351.18 357.18##
## Call: glm(formula = Industry ~ Milk + Frozen + Detergents_Paper, family = "binomial",
## data = wholesale_train)
##
## Coefficients:
## (Intercept) Milk Frozen Detergents_Paper
## -0.4646 0.9042 -0.8845 4.6254
##
## Degrees of Freedom: 439 Total (i.e. Null); 436 Residual
## Null Deviance: 553.4
## Residual Deviance: 208.4 AIC: 216.4wholesale.logit <- glm(
formula = Industry ~ Milk + Frozen + Detergents_Paper,
family = "binomial", data = wholesale_train
)## Warning: glm.fit: fitted probabilities numerically 0 or 1 occurredAfter that, let’s check again the multicollinearity using vif function.
p_load(car)
vif(wholesale.logit)## Milk Frozen Detergents_Paper
## 1.182963 1.135161 1.077362It results small values. So, let’s predict the test set.
wholesale_test$pred.Risk <- predict(wholesale.logit, wholesale_test, type = "response")Create accuracy function.
accuracy_log <- function(data) {
# initialize number of predictions
correct <- 0
for (i in c(1:nrow(data))) {
# 7th variable is actual class, 8th is our predicted class
if (data[i, 10] == data[i, 11]) {
correct <- correct + 1
}
}
percentage <- correct / nrow(data) * 100
return(percentage)
}And finally, we find the accuracy.
wholesale_test$pred_log <- as.numeric(wholesale_test$pred.Risk >= 0.3)
wholesale_test$actual <- as.numeric(wholesale_test$Industry == "retail")
accuracy_log(wholesale_test)## [1] 91.36364The accuracy of logistic regression (91.36) is slighly better the accuracy of KNN (90.91). It’s likely happen after we remove some variable to predict with logistic regression. The accuracy probably results differently, if we also use only some variables to predict with KNN.
Using KNN, we don’t need to assure the assumption, such as no multicollinearity among variables, and linearity of variables. In logistic regression process, we need to see the variable correlation first, and using step-wise regression, also VIF.
Compare to KNN, logistic regression don’t need to scaling the data, and it also works if there are non-numeric variables.