findKnee.R

findKnee.R

findknee <- function(xdata)
{
  rate_of_change=(xdata[1]-xdata[length(xdata)])/(length(xdata)-1)
  xdata$delta = xdata-xdata[length(xdata)]
  xdata$deltas[1] = xdata$delta[1]
  for (d in 2:length(xdata))
  {
    xdata$deltas[d]=xdata$deltas[d-1]-rate_of_change
  }
  for (d in 1:length(xdata))
  {
    xdata$deltas[d]=xdata$delta[d]-xdata$deltas[d]
  }
  return(abs(xdata$deltas))
}

#library("factoextra")
#finds the convex or invex of any given curve by subtracting the min/max's linear plane and then finding the absolute max.

df <- read.csv(file="C:\\Users\\User\\Documents\\wiki\\wiki\\dev\\python\\python-ml\\data\\raw\\states.csv",text=readings, header = TRUE, sep = ",", dec = ".")

data <- df[,2:ncol(df)]

data_scaled <- scale(data)

#f_data <- fviz_nbclust(data_scaled, kmeans, method = "wss", k.max = 24) + theme_minimal() + ggtitle("the Elbow Method")

set.seed(31)

metric1 <- c()
metric2 <- c()
for (k in 1:15)
{#k=2
  #print(k)
  km <- kmeans(data_scaled, nstart=100, centers=k)
  metric1 <- c(metric1,km$betweenss)
  metric2 <- c(metric2,sum(km$withinss)/length(km$withinss))
}

plot(metric1)
plot(metric2)

knee1 = (findknee(metric1))
knee1 = knee1[complete.cases(knee1)]
knee1 = knee1/max(knee1)
knee2 = (findknee(metric2))
knee2 = knee2[complete.cases(knee2)]
knee2 = knee2/max(knee2)

print(which.max(knee1))
print(which.max(knee2))

plot(knee1)
lines(knee1)
lines(knee2)

optimal_k <- which.min(abs(knee1/knee2-1))

Idea behind it

which(knee1/knee2==1)
[1] 4
print(which.max(knee1))
[1] 4
print(which.max(knee2))
[1] 4

def findknee(xdata):
    rate_of_change=(xdata[0]-xdata[-1])/(len(xdata)-1)
    #print(rate_of_change)
    delta = xdata-xdata[-1]
    deltas = []
    deltas.append(delta[0])
    for d in range(1,len(xdata)):
        deltas.append(deltas[d-1]-rate_of_change)
    #print(deltas)
    for d in range(0,len(xdata)):
        deltas[d]=delta[d]-deltas[d]
    return(np.round(np.abs(deltas)))