kmean_samesize: kmean_samesize

In 1Edtrujillo1/udeploy: Back-end & Front-end functions

Description

Create optimal k-mean labels for a dataset.

Usage

kmean_samesize(df, k, nstart = 25)

Arguments

df	dataset to generate k-means labels
k	number of desire clusters
nstart	number of k-mean iterations for the centroids creation.

Details

1. Calculate the k-means algorithm on numeric or integar variables from the dataset df.

2. Calculate the euclidean distance between the row df points and each centroid.

• (x0,y0,z0,..),(x1,y1,z1,...),..,(xk,yk,zk,...) k centroids.

• We are going to have k df s. Therefore, we use cbind.data.frame to create a df where each df is a variable with length of the original df.

1. 3.variables used:

• cardinality_sample: number of elements that are going to be for each cluster (label).

• ctrl_clstr_no_elmnts: this variable helps to control the balance cluster size.

• label: result balanced label that is going to be created.

1. 4.For each iteration (each df row):

2. 4.1. We obtain the row index where is the minimum euclidean distance in that row of the df 2. defined as bestcluster and in that index we add 1 to the observation of ctrl_clstr_no_elmnts variable representing adding 1 intended top to k.

• example:

• iteration 1 - bestcluster = 2

• ctrl_clstr_no_elmnts = 0 1 0 0 0 0 | row 1 of dataset euclidean_distance

• iteration 2 - bestcluster = 2

• ctrl_clstr_no_elmnts = 0 2 0 0 0 0 | row 2 of dataset euclidean_distance

• iteration 3 - bestcluster = 3

• ctrl_clstr_no_elmnts = 0 2 1 0 0 0 | row 3 of dataset euclidean_distance

• iteration 4 - bestcluster = 2

• ctrl_clstr_no_elmnts = 0 3 1 0 0 0 | row 4 of dataset euclidean_distance

1. 4.2.Add if statement to control and balance the size of each cluster based on ctrl_clstr_no_elmnts variable. If in the index bestcluster of the observation of ctrl_clstr_no_elmnts is greater than cardinality_sample then define the variable on index bestcluster of df2. as NA to not affect the next iteration and have only the columns in euclidean_distance where ctrl_clstr_no_elmnts is less than cardinality_sample

Value

balanced k-means label for each observation of the dataset

Note

• The kmeans algorithm assigns a label for each observation. However, it is not necessary to have balanced clusters in the amount of observations that is why we use this function.

• Kmeans algorithm assign randomly the centroids based on the iterations of nstart. This is why every time this function runs the output would have a different order in the labels.

• It is recommendable that the size of the population is divisible by the number of clusters k to have equal number of elements in each cluster.

Author(s)

Eduardo Trujillo

1Edtrujillo1/udeploy documentation built on July 13, 2021, 9:12 p.m.