I found data base “Wholesale customers data.csv” on the website www.kaggle.com. Link of the data set: https://www.kaggle.com/datasets/binovi/wholesale-customers-data-set . In this data base there is data about 440 observations.
The data set refers to customers of a wholesale distributor. It includes the annual spending in monetary units on diverse product categories.
The research question: Can we cluster customers of a wholesale distributor based on their annual spending patterns in monetary units to identify homogeneous groups of customers with similar purchasing behaviors and preferences?
Since the research questions is whether we can group customers, the best method is using clustering.
mydata <- read.csv("Wholesale customers data.csv")
# Create a new column 'ID' with numbering starting from 1
mydata$ID <- seq_len(nrow(mydata))
First I added one more column to the dataset representing ID. Unit is one customer and sample size is 440 customers.
head(mydata)
## Channel Region Fresh Milk Grocery Frozen Detergents_Paper Delicassen ID
## 1 2 3 12669 9656 7561 214 2674 1338 1
## 2 2 3 7057 9810 9568 1762 3293 1776 2
## 3 2 3 6353 8808 7684 2405 3516 7844 3
## 4 1 3 13265 1196 4221 6404 507 1788 4
## 5 2 3 22615 5410 7198 3915 1777 5185 5
## 6 2 3 9413 8259 5126 666 1795 1451 6
Description:
describe(mydata[,-9])
## mydata[, -9]
##
## 8 Variables 440 Observations
## --------------------------------------------------------------------------------
## Channel
## n missing distinct Info Mean Gmd
## 440 0 2 0.656 1.323 0.4381
##
## Value 1 2
## Frequency 298 142
## Proportion 0.677 0.323
## --------------------------------------------------------------------------------
## Region
## n missing distinct Info Mean Gmd
## 440 0 3 0.623 2.543 0.6951
##
## Value 1 2 3
## Frequency 77 47 316
## Proportion 0.175 0.107 0.718
##
## For the frequency table, variable is rounded to the nearest 0
## --------------------------------------------------------------------------------
## Fresh
## n missing distinct Info Mean Gmd .05 .10
## 440 0 433 1 12000 12256 401.9 915.6
## .25 .50 .75 .90 .95
## 3127.8 8504.0 16933.8 27090.5 36818.5
##
## lowest : 3 9 18 23 37, highest: 56083 56159 68951 76237 112151
## --------------------------------------------------------------------------------
## Milk
## n missing distinct Info Mean Gmd .05 .10
## 440 0 421 1 5796 6152 593.8 889.8
## .25 .50 .75 .90 .95
## 1533.0 3627.0 7190.2 12229.9 16843.4
##
## lowest : 55 112 134 201 254, highest: 38369 43950 46197 54259 73498
## --------------------------------------------------------------------------------
## Grocery
## n missing distinct Info Mean Gmd .05 .10
## 440 0 430 1 7951 8365 851.5 1381.9
## .25 .50 .75 .90 .95
## 2153.0 4755.5 10655.8 18910.1 24033.5
##
## lowest : 3 137 218 223 245, highest: 45828 55571 59598 67298 92780
## --------------------------------------------------------------------------------
## Frozen
## n missing distinct Info Mean Gmd .05 .10
## 440 0 426 1 3072 3635 136.8 281.3
## .25 .50 .75 .90 .95
## 742.2 1526.0 3554.2 7545.3 9930.7
##
## lowest : 25 33 36 38 42, highest: 18028 18711 35009 36534 60869
## --------------------------------------------------------------------------------
## Detergents_Paper
## n missing distinct Info Mean Gmd .05 .10
## 440 0 417 1 2881 3927 63.7 99.6
## .25 .50 .75 .90 .95
## 256.8 816.5 3922.0 7438.3 12043.2
##
## lowest : 3 5 7 9 10, highest: 24171 24231 26701 38102 40827
## --------------------------------------------------------------------------------
## Delicassen
## n missing distinct Info Mean Gmd .05 .10
## 440 0 403 1 1525 1677 63.95 180.80
## .25 .50 .75 .90 .95
## 408.25 965.50 1820.25 2945.90 4485.40
##
## lowest : 3 7 8 11 18, highest: 8550 14351 14472 16523 47943
## --------------------------------------------------------------------------------
Now I did some descriptive statistics. I put mydata[,-9], because there is no point in analyzing ID.
We can see that there are no missing values in any of the variables (that is good).
We can see that there are 298 customers with channel 1 (Hotel/Restaurant/Cafee) and 142 customers with channel 2 (Retail).
The mean for Fresh is 12000 m.u., that means that on average, the customers of the wholesale distributor spend 12000 m.u. per year on fresh products.
On the other hand, the mean for Milk is 5796 m.u., that means that on average, the customers of the wholesale distributor spend 5796 m.u. per year on milk products.
The customers of the wholesale distributor spend on average 7951 m.u. per year on grocery products.
The customers of the wholesale distributor spend on average 2881 m.u. per year on on detergents and paper products.
summary(mydata[c("Fresh", "Milk", "Grocery", "Detergents_Paper")]) #Describing clustering variables
## Fresh Milk Grocery Detergents_Paper
## Min. : 3 Min. : 55 Min. : 3 Min. : 3.0
## 1st Qu.: 3128 1st Qu.: 1533 1st Qu.: 2153 1st Qu.: 256.8
## Median : 8504 Median : 3627 Median : 4756 Median : 816.5
## Mean : 12000 Mean : 5796 Mean : 7951 Mean : 2881.5
## 3rd Qu.: 16934 3rd Qu.: 7190 3rd Qu.:10656 3rd Qu.: 3922.0
## Max. :112151 Max. :73498 Max. :92780 Max. :40827.0
50% of the customers spend less or equal to 8504 m.u. per year on fresh products, while the other 50% spend more or equal to 8504 m.u. 50% of the customers spend less or equal to 3627 m.u. per year on milk products, while the other 50% spend more or equal to 3627 m.u. 50% of the customers spend less or equal to 4756 m.u. per year on grocery products, while the other 50% spend more or equal to 4756 m.u. 50% of the customers spend less or equal to 816.5 m.u. per year on detergents and paper products, while the other 50% spend more or equal to 816.5 m.u.
#In order to do clustering, we need to first standardize all 4 cluster variables
mydata$Fresh_z <- scale(mydata$Fresh)
mydata$Milk_z <- scale(mydata$Milk)
mydata$Grocery_z <- scale(mydata$Grocery)
mydata$Detergents_Paper_z <- scale(mydata$Detergents_Paper)
For this research I took 4 cluster variables: Fresh, Milk, Grocery and Detergents_Paper. Since all 4 variables have different range of values, we standardize them. Now they are comparable.
library(Hmisc)
rcorr(as.matrix(mydata[, c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")]), type = "pearson")
## Fresh_z Milk_z Grocery_z Detergents_Paper_z
## Fresh_z 1.00 0.10 -0.01 -0.10
## Milk_z 0.10 1.00 0.73 0.66
## Grocery_z -0.01 0.73 1.00 0.92
## Detergents_Paper_z -0.10 0.66 0.92 1.00
##
## n= 440
##
##
## P
## Fresh_z Milk_z Grocery_z Detergents_Paper_z
## Fresh_z 0.0351 0.8042 0.0325
## Milk_z 0.0351 0.0000 0.0000
## Grocery_z 0.8042 0.0000 0.0000
## Detergents_Paper_z 0.0325 0.0000 0.0000
With this step we are chacking if cluster variables correlate.
mydata$Dissimilarity <- sqrt(mydata$Fresh_z^2 + mydata$Milk_z^2
+ mydata$Grocery_z^2 + mydata$Detergents_Paper_z^2)
#Finding outliers
head(mydata[order(-mydata$Dissimilarity), ], 10) #10 units with the highest value of dissimilarity
## Channel Region Fresh Milk Grocery Frozen Detergents_Paper Delicassen ID
## 86 2 3 16117 46197 92780 1026 40827 2944 86
## 87 2 3 22925 73498 32114 987 20070 903 87
## 48 2 3 44466 54259 55571 7782 24171 6465 48
## 334 2 2 8565 4980 67298 131 38102 1215 334
## 62 2 3 35942 38369 59598 3254 26701 2017 62
## 182 1 3 112151 29627 18148 16745 4948 8550 182
## 66 2 3 85 20959 45828 36 24231 1423 66
## 184 1 3 36847 43950 20170 36534 239 47943 184
## 212 2 1 12119 28326 39694 4736 19410 2870 212
## 93 2 3 9198 27472 32034 3232 18906 5130 93
## Fresh_z Milk_z Grocery_z Detergents_Paper_z Dissimilarity
## 86 0.325499741 5.4740744 8.926367 7.9586127 13.156405
## 87 0.863795224 9.1732079 2.542598 3.6050821 10.215445
## 48 2.567000718 6.5664305 5.010934 4.4652174 9.734225
## 334 -0.271622393 -0.1105995 6.244944 7.3870768 9.677512
## 62 1.893024411 4.4134240 5.434688 4.9958544 8.806602
## 182 7.918723656 3.2289317 1.072982 0.4334249 8.629677
## 66 -0.942119704 2.0544660 3.985696 4.4778017 6.406630
## 184 1.964581022 5.1696185 1.285753 -0.5542311 5.704811
## 212 0.009385561 3.0526535 3.340227 3.4666551 5.700315
## 93 -0.221572299 2.9369412 2.534180 3.3609471 5.137388
#most dissimilar - furthest away from the center
We can see that customer with ID 86 is the most dissimilar.
library(factoextra)
#Calculating Euclidean distances
distance <- get_dist(mydata[c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")],
method = "euclidian")
distance2 <- distance^2
fviz_dist(distance2) #Showing dissimilarity matrix
get_clust_tendency(mydata[c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")], #Hopkins statistics
n = nrow(mydata) - 1,
graph = FALSE)
## $hopkins_stat
## [1] 0.9330562
##
## $plot
## NULL
Here we are checking if data is clustarable, if Hopkins Statistics is less then 0.5 that would mean data is not appropriate. But our Hopkins Statistics is very high so we have nothing to worry about.
library(dplyr) #Allowing use of %>%
WARD <- mydata[c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")] %>%
#scale() %>%
get_dist(method = "euclidean") %>% #Selecting distance
hclust(method = "ward.D2") #Selecting algorithm
WARD
##
## Call:
## hclust(d = ., method = "ward.D2")
##
## Cluster method : ward.D2
## Distance : euclidean
## Number of objects: 440
library(factoextra)
fviz_dend(WARD)
## Warning: The `<scale>` argument of `guides()` cannot be `FALSE`. Use "none" instead as
## of ggplot2 3.3.4.
## ℹ The deprecated feature was likely used in the factoextra package.
## Please report the issue at <https://github.com/kassambara/factoextra/issues>.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
According to the dendrogram, it would make most sense to make 3
groups.
fviz_dend(WARD,
k = 2,
cex = 0.5,
palette = "jama",
color_labels_by_k = TRUE,
rect = TRUE)
fviz_dend(WARD,
k = 3,
cex = 0.5,
palette = "jama",
color_labels_by_k = TRUE,
rect = TRUE)
fviz_dend(WARD,
k = 4,
cex = 0.5,
palette = "jama",
color_labels_by_k = TRUE,
rect = TRUE)
fviz_dend(WARD,
k = 5,
cex = 0.5,
palette = "jama",
color_labels_by_k = TRUE,
rect = TRUE)
set.seed(1)
install.packages("NbClust")
## Warning: package 'NbClust' is in use and will not be installed
library(NbClust)
OptNumber <- mydata[c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")] %>%
#scale() %>%
NbClust(distance = "euclidean",
min.nc = 2, max.nc = 10,
method = "ward.D2",
index = "all")
## *** : The Hubert index is a graphical method of determining the number of clusters.
## In the plot of Hubert index, we seek a significant knee that corresponds to a
## significant increase of the value of the measure i.e the significant peak in Hubert
## index second differences plot.
##
## *** : The D index is a graphical method of determining the number of clusters.
## In the plot of D index, we seek a significant knee (the significant peak in Dindex
## second differences plot) that corresponds to a significant increase of the value of
## the measure.
##
## *******************************************************************
## * Among all indices:
## * 2 proposed 2 as the best number of clusters
## * 9 proposed 3 as the best number of clusters
## * 2 proposed 4 as the best number of clusters
## * 4 proposed 6 as the best number of clusters
## * 2 proposed 7 as the best number of clusters
## * 1 proposed 8 as the best number of clusters
## * 3 proposed 10 as the best number of clusters
##
## ***** Conclusion *****
##
## * According to the majority rule, the best number of clusters is 3
##
##
## *******************************************************************
As we saw on the dandogram, here we comfirm that the best number of clusters is 3.
mydata$ClusterWard <- cutree(WARD,
k = 3) #Cutting the dendrogram
head(mydata)
## Channel Region Fresh Milk Grocery Frozen Detergents_Paper Delicassen ID
## 1 2 3 12669 9656 7561 214 2674 1338 1
## 2 2 3 7057 9810 9568 1762 3293 1776 2
## 3 2 3 6353 8808 7684 2405 3516 7844 3
## 4 1 3 13265 1196 4221 6404 507 1788 4
## 5 2 3 22615 5410 7198 3915 1777 5185 5
## 6 2 3 9413 8259 5126 666 1795 1451 6
## Fresh_z Milk_z Grocery_z Detergents_Paper_z Dissimilarity
## 1 0.05287300 0.52297247 -0.04106815 -0.04351919 0.5290333
## 2 -0.39085706 0.54383861 0.17012470 0.08630859 0.6963629
## 3 -0.44652098 0.40807319 -0.02812509 0.13308016 0.6200049
## 4 0.09999758 -0.62331041 -0.39253008 -0.49802132 0.8947740
## 5 0.83928412 -0.05233688 -0.07926595 -0.23165413 0.8758332
## 6 -0.20457266 0.33368675 -0.29729863 -0.22787885 0.5417675
## ClusterWard
## 1 1
## 2 1
## 3 1
## 4 2
## 5 2
## 6 2
#Calculating positions of initial leaders
Initial_leaders <- aggregate(mydata[, c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")],
by = list(mydata$ClusterWard),
FUN = mean)
Initial_leaders
## Group.1 Fresh_z Milk_z Grocery_z Detergents_Paper_z
## 1 1 -0.5205590 0.7529972 0.8919869 0.8981549
## 2 2 0.1702706 -0.3503476 -0.4085563 -0.4115614
## 3 3 1.0755395 5.1033075 5.6319063 5.6823687
We have 3 initial leaders because we have 3 groups (clusters). Each leader is defined with 4 means from cluster variables.
Centroid 1 = c(-0.5205590, 0.7529972, 0.8919869, 0.8981549)
library(factoextra)
#Performing K-Means clustering - initial leaders are chosen as centroids of groups, found with hierarhical clustering
K_MEANS <- hkmeans(mydata[c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")],
k = 3, #number of clusters previously determend
hc.metric = "euclidean",
hc.method = "ward.D2")
K_MEANS
## Hierarchical K-means clustering with 3 clusters of sizes 102, 328, 10
##
## Cluster means:
## Fresh_z Milk_z Grocery_z Detergents_Paper_z
## 1 -0.34182554 0.7461291 0.9075902 0.8554505
## 2 0.09674229 -0.3514623 -0.4124442 -0.4066619
## 3 0.31347349 3.9174467 4.2707490 4.6129149
##
## Clustering vector:
## [1] 2 1 2 2 2 2 2 2 2 1 1 2 1 1 1 2 1 2 2 2 2 2 2 1 1 2 2 2 1 2 2 2 2 2 2 1 2
## [38] 1 1 2 2 2 1 1 1 1 1 3 1 1 2 2 2 1 2 2 3 1 2 2 2 3 2 1 2 3 2 1 2 2 2 1 2 2
## [75] 2 2 2 1 2 2 2 1 1 2 2 3 3 2 2 2 2 2 3 2 1 2 2 2 2 2 1 1 2 2 2 2 1 1 2 1 2
## [112] 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 2 2
## [149] 2 2 2 2 2 2 2 1 1 2 1 1 1 2 2 1 2 1 1 2 2 2 1 1 2 1 2 1 2 2 2 2 2 1 1 1 2
## [186] 2 2 1 1 1 2 2 2 1 2 2 2 1 2 2 1 1 2 2 2 1 2 2 2 1 2 3 2 2 1 1 1 2 1 2 2 1
## [223] 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 1 2 2 2 2 2 3 2 2 1 2 2 2 2
## [260] 2 2 2 2 2 1 1 1 2 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 1 2 2
## [297] 2 2 2 2 2 1 1 1 1 1 1 2 2 1 2 2 1 2 2 1 2 2 2 1 2 2 2 2 2 1 2 2 2 2 2 1 2
## [334] 3 2 2 2 2 2 2 1 1 2 1 2 2 1 2 2 1 2 1 2 1 2 2 2 1 2 2 2 2 2 2 2 2 2 2 2 2
## [371] 2 2 2 2 2 2 1 2 2 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 2
## [408] 1 2 2 2 2 2 2 2 2 1 2 1 2 1 2 2 2 2 2 1 2 2 2 2 2 2 2 2 2 2 1 2 2
##
## Within cluster sum of squares by cluster:
## [1] 283.2396 399.8343 142.3068
## (between_SS / total_SS = 53.0 %)
##
## Available components:
##
## [1] "cluster" "centers" "totss" "withinss" "tot.withinss"
## [6] "betweenss" "size" "iter" "ifault" "data"
## [11] "hclust"
Now values of leaders have changed because some customers we reclassified. Ratio between sum of squares and total sum of squares is 53%, which is okey. (If it was less than 50% that would be bad)
fviz_cluster(K_MEANS,
palette = "jama",
repel = FALSE,
ggtheme = theme_classic())
Now this is a representation of clusters but it is not in 2D so we can
ignore the “overlapping” between groups. But we see that customer ID 182
differs from other customers in his group. Because we want the groups to
be as homogeneous as possible, we will remove this customer.
mydata <- mydata %>%
filter(!ID %in% c(182))
Now sample size is 439 since we removed one customer.
mydata_clu_std <- as.data.frame(scale(mydata[c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z")]))
library(factoextra)
K_MEANS1 <- hkmeans(mydata_clu_std, #Data
k = 3, #Number of groups
hc.metric = "euclidean", #Distance for hierar. clus.
hc.method = "ward.D2") #Algorithm for hierar. clus.
K_MEANS1
## Hierarchical K-means clustering with 3 clusters of sizes 323, 106, 10
##
## Cluster means:
## Fresh_z Milk_z Grocery_z Detergents_Paper_z
## 1 0.1517070 -0.3528215 -0.4138333 -0.4139944
## 2 -0.4960250 0.7007826 0.8578182 0.8266401
## 3 0.3577294 3.9678379 4.2739441 4.6096328
##
## Clustering vector:
## [1] 1 2 2 1 1 1 1 1 1 2 2 1 1 2 2 1 2 1 1 1 1 1 1 2 2 1 1 1 2 1 1 1 1 1 1 2 1
## [38] 2 2 1 1 1 2 2 2 2 2 3 2 2 1 1 1 2 1 1 3 2 1 1 1 3 1 2 1 3 1 2 1 1 1 2 1 1
## [75] 1 1 1 2 1 1 1 2 2 1 1 3 3 1 1 1 1 1 3 1 2 1 1 1 1 1 2 2 2 1 1 1 2 2 1 2 1
## [112] 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1
## [149] 1 1 1 1 1 1 1 2 2 1 2 2 2 1 1 2 2 2 2 1 1 1 2 2 1 2 1 2 1 1 1 1 1 2 2 1 1
## [186] 1 2 2 2 1 1 1 2 1 1 1 2 1 1 2 2 1 1 1 2 1 1 1 2 1 3 1 1 2 2 2 1 2 1 1 2 1
## [223] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 1 1 1 1 1 3 1 1 2 1 1 1 1 1
## [260] 1 1 1 1 2 2 2 1 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1
## [297] 1 1 1 1 2 2 2 2 2 2 1 1 2 1 1 2 1 1 2 1 1 1 2 1 1 1 1 1 1 1 1 1 1 1 2 1 3
## [334] 1 1 1 1 1 1 2 2 1 2 1 1 2 1 1 2 1 2 1 2 1 1 1 2 2 1 1 1 1 1 1 2 1 1 1 1 1
## [371] 1 1 1 1 1 2 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 1 1 1 2
## [408] 1 1 1 1 2 1 1 1 2 2 2 1 2 1 1 1 1 1 2 1 1 1 1 1 1 1 1 1 1 2 1 1
##
## Within cluster sum of squares by cluster:
## [1] 447.8816 217.8134 145.5470
## (between_SS / total_SS = 53.7 %)
##
## Available components:
##
## [1] "cluster" "centers" "totss" "withinss" "tot.withinss"
## [6] "betweenss" "size" "iter" "ifault" "data"
## [11] "hclust"
Now values of leaders have changed because some customers we reclassified. Ratio between sum of squares and total sum of squares is 53.7%, which is also okey.
library(factoextra)
fviz_cluster(K_MEANS1,
palette = "jama",
repel = FALSE,
ggtheme = theme_classic())
Now the groups look much more homogeneous.
mydata$ClusterK_Means <- K_MEANS1$cluster
head(mydata[c("ID", "ClusterWard", "ClusterK_Means")])
## ID ClusterWard ClusterK_Means
## 1 1 1 1
## 2 2 1 2
## 3 3 1 2
## 4 4 2 1
## 5 5 2 1
## 6 6 2 1
Here we can see that both customers with ID 2 and ID 3 ware in group 1 yet they got reclassified to group 2. Whereas customers with ID 4, ID 5 and ID 6 ware in group 2 yet they got reclassified to group 1.
#Checking for reclassifications
table(mydata$ClusterWard)
##
## 1 2 3
## 115 319 5
table(mydata$ClusterK_Means)
##
## 1 2 3
## 323 106 10
table(mydata$ClusterWard, mydata$ClusterK_Means)
##
## 1 2 3
## 1 9 101 5
## 2 314 5 0
## 3 0 0 5
Group 1 initially included 115 customers, 9 customers stayed in group 1, 101 customers were reclassified to group 2 and 5 were reclassified to group 3.
Group 1 currently has 323 customers, 9 were already there and 314 reclassified from group 2 to group 1.
#final leaders
Centroids <- K_MEANS1$centers
Centroids
## Fresh_z Milk_z Grocery_z Detergents_Paper_z
## 1 0.1517070 -0.3528215 -0.4138333 -0.4139944
## 2 -0.4960250 0.7007826 0.8578182 0.8266401
## 3 0.3577294 3.9678379 4.2739441 4.6096328
library(ggplot2)
library(tidyr)
## Warning: package 'tidyr' was built under R version 4.3.2
##
## Attaching package: 'tidyr'
## The following object is masked from 'package:magrittr':
##
## extract
Figure <- as.data.frame(Centroids)
Figure$id <- 1:nrow(Figure)
Figure <- pivot_longer(Figure, cols = c(Fresh_z, Milk_z, Grocery_z, Detergents_Paper_z))
Figure$Groups <- factor(Figure$id,
levels = c(1, 2, 3),
labels = c("1", "2", "3"))
Figure$nameFactor <- factor(Figure$name,
levels = c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z"),
labels = c("Fresh_z", "Milk_z", "Grocery_z", "Detergents_Paper_z"))
ggplot(Figure, aes(x = nameFactor, y = value)) +
geom_hline(yintercept = 0) +
theme_bw() +
geom_point(aes(shape = Groups, col = Groups), size = 3) +
geom_line(aes(group = id), linewidth = 1) +
ylab("Averages") +
xlab("Cluster variables") +
ylim(-2, 5)
We can see that group 3 spends much more on milk, grocery and detergents
and paper products compared to other two groups. When it comes to fresh
products groups do not differ as much.
Group 1 spends more on fresh products compared to other products. Group 2 spends less on fresh products compared to other products.
#Are all the cluster variables successful at classifying units into groups? Performing ANOVAs.
fit <- aov(cbind(Fresh_z, Milk_z, Grocery_z, Detergents_Paper_z) ~ as.factor(ClusterK_Means),
data = mydata)
summary(fit)
## Response 1 :
## Df Sum Sq Mean Sq F value Pr(>F)
## as.factor(ClusterK_Means) 2 29.88 14.9404 18.812 1.457e-08 ***
## Residuals 436 346.27 0.7942
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Response 2 :
## Df Sum Sq Mean Sq F value Pr(>F)
## as.factor(ClusterK_Means) 2 244.31 122.157 289.09 < 2.2e-16 ***
## Residuals 436 184.24 0.423
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Response 3 :
## Df Sum Sq Mean Sq F value Pr(>F)
## as.factor(ClusterK_Means) 2 315.87 157.94 564.54 < 2.2e-16 ***
## Residuals 436 121.97 0.28
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Response 4 :
## Df Sum Sq Mean Sq F value Pr(>F)
## as.factor(ClusterK_Means) 2 340.91 170.455 759.12 < 2.2e-16 ***
## Residuals 436 97.90 0.225
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
For response 1: H0: Mu fresh(G1) = Mu fresh(G2) = Mu fresh(G3) H1: At least one is different.
We reject H0 at p-value < 0.001. (If We didn’t reject H0 that would mean there is overlapping)
For response 2: H0: Mu milk(G1) = Mu milk(G2) = Mu milk(G3) H1: At least one is different.
We reject H0 at p-value < 0.001.
For response 3: H0: Mu grocery(G1) = Mu grocery(G2) = Mu grocery(G3) H1: At least one is different.
We reject H0 at p-value < 0.001.
For response 4: H0: Mu detergents_paper(G1) = Mu detergents_paper(G2) = Mu detergents_paper(G3) H1: At least one is different.
We reject H0 at p-value < 0.001.
mydata$Region_F <- factor(mydata$Region,
levels = c(1, 2, 3),
labels = c("1", "2", "3"))
#Pearson Chi2 test
chisq_results <- chisq.test(mydata$Region_F, as.factor(mydata$ClusterK_Means))
## Warning in chisq.test(mydata$Region_F, as.factor(mydata$ClusterK_Means)):
## Chi-squared approximation may be incorrect
chisq_results
##
## Pearson's Chi-squared test
##
## data: mydata$Region_F and as.factor(mydata$ClusterK_Means)
## X-squared = 0.27312, df = 4, p-value = 0.9915
H0: There is no association between Region and Cluster variables. H1: There is association between Region and Cluster variables.
We can not reject H0 at p-value = 0.992. There is no association between Region and Cluster variables.
mydata$Channel_F <- factor(mydata$Channel,
levels = c(1, 2),
labels = c("1", "2"))
chisq_results <- chisq.test(mydata$Channel_F, as.factor(mydata$ClusterK_Means))
## Warning in chisq.test(mydata$Channel_F, as.factor(mydata$ClusterK_Means)):
## Chi-squared approximation may be incorrect
chisq_results
##
## Pearson's Chi-squared test
##
## data: mydata$Channel_F and as.factor(mydata$ClusterK_Means)
## X-squared = 216.58, df = 2, p-value < 2.2e-16
H0: There is no association between Channel and Cluster variables. H1: There is association between Channel and Cluster variables.
We reject H0 at p-value < 0.001. There is association between Channel and Cluster variables.
addmargins(chisq_results$observed)
##
## mydata$Channel_F 1 2 3 Sum
## 1 282 15 0 297
## 2 41 91 10 142
## Sum 323 106 10 439
round(chisq_results$expected, 2)
##
## mydata$Channel_F 1 2 3
## 1 218.52 71.71 6.77
## 2 104.48 34.29 3.23
More that 20% of values are above 5 so the assumtion is met. (Otherwise we would need to do Fisher Test of Indep.).
round(chisq_results$res, 2)
##
## mydata$Channel_F 1 2 3
## 1 4.29 -6.70 -2.60
## 2 -6.21 9.69 3.76
4.29 - More than expected number of customers are in channel 1 (Hotel/Restaurant/Cafee) at alpha = 0.1
aggregate(mydata$Frozen,
by = list(mydata$ClusterK_Means),
FUN = "mean")
## Group.1 x
## 1 1 3409.910
## 2 2 1914.698
## 3 3 3054.600
fit <- aov(Frozen ~ as.factor(ClusterK_Means),
data = mydata)
summary(fit)
## Df Sum Sq Mean Sq F value Pr(>F)
## as.factor(ClusterK_Means) 2 1.784e+08 89213681 3.897 0.021 *
## Residuals 436 9.980e+09 22891013
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
H0: Mu frozen(G1) = Mu frozen(G2) = Mu frozen(G3) H1: At least one is different.
We reject H0 at p-value = 0.021. (If We didn’t reject H0 that would mean there is overlapping)
aggregate(mydata$Delicassen,
by = list(mydata$ClusterK_Means),
FUN = "mean")
## Group.1 x
## 1 1 1191.412
## 2 2 2340.934
## 3 3 2942.800
fit <- aov(Delicassen ~ as.factor(ClusterK_Means),
data = mydata)
summary(fit)
## Df Sum Sq Mean Sq F value Pr(>F)
## as.factor(ClusterK_Means) 2 1.265e+08 63250216 8.318 0.000285 ***
## Residuals 436 3.315e+09 7604130
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
H0: Mu delicassen(G1) = Mu delicassen(G2) = Mu delicassen(G3) H1: At least one is different.
We reject H0 at p-value = 0.001. (If We didn’t reject H0 that would mean there is overlapping)
RESULTS
We divided 439 customers of a wholesale distributor into 3 groups based on four standardized variables.
For hierarchical clustering, we used Ward’s clustering algorithm and decided to divide them into 3 groups based on the analysis of the dendrogram. The classification was further optimized using K-Means clustering.
Group 1 contains the most customers (73.6% - 323/439), characterized by higher than average annual spending on fresh products and lower than average spending on milk, grocery and detergents and paper products. There are more than expected number of customers from channel 1 (Hotel/Restaurant/Cafee) and less than expected number of customers with channel 2 (Retail) (alpha = 0.1). The customers in this group spend on average 3409.910 m.u. on frozen products and 1191.412 m.u. on delicassen products.