Problem
I have a dataframe that composes of > 5 variables at any time and am trying to do a K-Means of it. Because K-Means is greatly affected by outliers, I've been trying to look for a few hours on how to calculate and remove multivariate outliers. Most examples demonstrated are with 2 variables.
Possible Solutions Explored
mvoutlier - Kind user here noted that mvoutlier may be what I need.
Another Outlier Detection Method - Poster here commented with a mix of R functions to generate an ordered list of outliers.
Issues thus Far
Regarding mvoutlier, I was unable to generate a result because it noted my dataset contained negatives and it could not work because of that. I'm not sure how to alter my data to only positive since I need negatives in the set I am working with.
Regarding Another Outlier Detection Method I was able to come up with a list of outliers, but am unsure how to exclude them from the current data set. Also, I do know that these calculations are done after K-Means, and thus I probably will apply the math prior to doing K-Means.
Minimal Verifiable Example
Unfortunately, the dataset I'm using is off-limits to be shown to anyone, so what you'll need is any random data set with more than 3 variables. The code below is code converted from the Another Outlier Detection Method post to work with my data. It should work dynamically if you have a random data set as well. But it should have enough data where cluster center amount should be okay with 5.
clusterAmount <- 5
cluster <- kmeans(dataFrame, centers = clusterAmount, nstart = 20)
centers <- cluster$centers[cluster$cluster, ]
distances <- sqrt(rowSums(clusterDataFrame - centers)^2)
m <- tapply(distances, cluster$cluster, mean)
d <- distances/(m[cluster$cluster])
# 1% outliers
outliers <- d[order(d, decreasing = TRUE)][1:(nrow(clusterDataFrame) * .01)]
Output: A list of outliers ordered by their distance away from the center they reside in I believe. The issue then is getting these results paired up to the respective rows in the data frame and removing them so I can start my K-Means procedure. (Note, while in the example I used K-Means prior to removing outliers, I'll make sure to take the necessary steps and remove outliers before K-Means upon solution).
Question
With Another Outlier Detection Method example in place, how do I pair the results with the information in my current data frame to exclude those rows before doing K-Means?
I don't know if this is exactly helpful but if your data is multivariate normal you may want to try out a Wilks (1963) based method. Wilks showed that the mahalanobis distances of multivariate normal data follow a Beta distribution. We can take advantage of this (iris Sepal data used as an example):
test.dat <- iris[,-c(1,2))]
Wilks.function <- function(dat){
n <- nrow(dat)
p <- ncol(dat)
# beta distribution
u <- n * mahalanobis(dat, center = colMeans(dat), cov = cov(dat))/(n-1)^2
w <- 1 - u
F.stat <- ((n-p-1)/p) * (1/w-1) # computing F statistic
p <- 1 - round( pf(F.stat, p, n-p-1), 3) # p value for each row
cbind(w, F.stat, p)
}
plot(test.dat,
col = "blue",
pch = c(15,16,17)[as.numeric(iris$Species)])
dat.rows <- Wilks.function(test.dat); head(dat.rows)
# w F.stat p
#[1,] 0.9888813 0.8264127 0.440
#[2,] 0.9907488 0.6863139 0.505
#[3,] 0.9869330 0.9731436 0.380
#[4,] 0.9847254 1.1400985 0.323
#[5,] 0.9843166 1.1710961 0.313
#[6,] 0.9740961 1.9545687 0.145
Then we can simply find which rows of our multivariate data are significantly different from the beta distribution.
outliers <- which(dat.rows[,"p"] < 0.05)
points(test.dat[outliers,],
col = "red",
pch = c(15,16,17)[as.numeric(iris$Species[outliers])])
Related
To understand my problem, you will need the whole dataset: https://pastebin.com/82paf0G8
Pre-processing: I had a list of orders and 696 unique item numbers, and wanted to cluster them, based on how frequent each pair of items are ordered together. I calculated for each pair of items, number of frequency of occurence within the same order. I.e the highest number of occurrence was 489 between two items. I then "calculated" the similarity/correlation, by: Frequency / "max frequency of all pairs" (489). Now I have the dataset that I have uploaded.
Similarity/correlation: I don't know if my similarity approach is the best in this case. I also tried with something called "Jaccard’s coefficient/index", but get almost same results.
The dataset: The dataset contains material numbers V1 and V2. and N is the correlation between the two material numbers between 0 - 1.
With help from another one, I managed to create a distance matrix and use the PAM clustering.
Why PAM clustering? A data scientist suggest this: You have more than 95% of pairs without information, this makes all these materials are at the same distance and a single cluster very dispersed. This problem can be solved using a PAM algorithm, but still you will have a very concentrated group. Another solution is to increase the weight of the distances other than one.
Problem 1: The matrix is only 567x567. I think for clustering I need the 696x696 full matrix, even though a lot of them are zeros. But i'm not sure.
Problem 2: Clustering does not do very well. I get very concentrated clusters. A lot of items are clustered in the first cluster. Also, according to how you verify PAM clusters, my clustering results are poor. Is it due to the similarity analysis? What else should I use? Is it due to the 95% of data being zeros? Should I change the zeros to something else?
The whole code and results:
#Suppose X is the dataset
df <- data.table(X)
ss <- dcast(rbind(df, df[, .(V1 = V2, V2 = V1, N)]), V1~V2, value.var = "N")[, -1]
ss <- ss/max(ss, na.rm = TRUE)
ss[is.na(ss)] <- 0
diag(ss) <- 1
Now using the PAM clustering
dd2 <- as.dist(1 - sqrt(ss))
pam2 <- pam(dd2, 4)
summary(as.factor(pam2$clustering))
But I get very concentrated clusters, as:
1 2 3 4
382 100 23 62
I'm not sure where you get the 696 number from. After you rbind, you have a dataframe with 567 unique values for V1 and V2, and then you perform the dcast, and end up with a matrix as expected 567 x 567. Clustering wise I see no issue with your clusters.
dim(df) # [1] 7659 3
test <- rbind(df, df[, .(V1 = V2, V2 = V1, N)])
dim(test) # [1] 15318 3
length(unique(test$V1)) # 567
length(unique(test$V2)) # 567
test2 <- dcast(test, V1~V2, value.var = "N")[,-1]
dim(test2) # [1] 567 567
#Mayo, forget what the data scientist said about PAM. Since you've mentioned this work is for a thesis. Then from an academic viewpoint, your current justification to why PAM is required, does not hold any merit. Essentially, you need to either prove or justify why PAM is a necessity for your case study. And given the nature of (continuous) variables in the dataset, V1, V2, N, I do not see the logic on why PAM is applicable here (like I mentioned in the comments, PAM works best for mixed variables).
Continuing further, See this post on correlation detection in R;
# Objective: Detect Highly Correlated variables, visualize them and remove them
data("mtcars")
my_data <- mtcars[, c(1,3,4,5,6,7)]
# print the first 6 rows
head(my_data, 6)
# compute correlation matrix using the cor()
res<- cor(my_data)
round(res, 2) # Unfortunately, the function cor() returns only the correlation coefficients between variables.
# Visualize the correlation
# install.packages("corrplot")
library(corrplot)
corrplot(res, type = "upper", order = "hclust",
tl.col = "black", tl.srt = 45)
# Positive correlations are displayed in blue and negative correlations in red color. Color intensity and the size of the circle are proportional to the correlation coefficients. In the right side of the correlogram, the legend color shows the correlation coefficients and the corresponding colors.
# tl.col (for text label color) and tl.srt (for text label string rotation) are used to change text colors and rotations.
#Apply correlation filter at 0.80,
#install.packages("caret", dependencies = TRUE)
library(caret)
highlyCor <- colnames(my_data)[findCorrelation(res, cutoff = 0.80, verbose = TRUE)]
# show highly correlated variables
highlyCor
[1] "disp" "mpg"
removeHighCor<- findCorrelation(res, cutoff = 0.80) # returns indices of highly correlated variables
# remove highly correlated variables from the dataset
my_data<- my_data[,-removeHighCor]
[1] 32 4
Hope this helps.
As an assignment I had to develop and algorithm and generate a samples for a given geometric distribution with PMF
Using the inverse transform method, I came up with the following expression for generating the values:
Where U represents a value, or n values depending on the size of the sample, drawn from a Unif(0,1) distribution and p is 0.3 as stated in the PMF above.
I have the algorithm, the implementation in R and I already generated QQ Plots to visually assess the adjustment of the empirical values to the theoretical ones (generated with R), i.e., if the generated sample follows indeed the geometric distribution.
Now I wanted to submit the generated sample to a goodness of fit test, namely the Chi-square, yet I'm having trouble doing this in R.
[I think this was moved a little hastily, in spite of your response to whuber's question, since I think before solving the 'how do I write this algorithm in R' problem, it's probably more important to deal with the 'what you're doing is not the best approach to your problem' issue (which certainly belongs where you posted it). Since it's here, I will deal with the 'doing it in R' aspect, but I would urge to you go back an ask about the second question (as a new post).]
Firstly the chi-square test is a little different depending on whether you test
H0: the data come from a geometric distribution with parameter p
or
H0: the data come from a geometric distribution with parameter 0.3
If you want the second, it's quite straightforward. First, with the geometric, if you want to use the chi-square approximation to the distribution of the test statistic, you will need to group adjacent cells in the tail. The 'usual' rule - much too conservative - suggests that you need an expected count in every bin of at least 5.
I'll assume you have a nice large sample size. In that case, you'll have many bins with substantial expected counts and you don't need to worry so much about keeping it so high, but you will still need to choose how you will bin the tail (whether you just choose a single cut-off above which all values are grouped, for example).
I'll proceed as if n were say 1000 (though if you're testing your geometric random number generation, that's pretty low).
First, compute your expected counts:
dgeom(0:20,.3)*1000
[1] 300.0000000 210.0000000 147.0000000 102.9000000 72.0300000 50.4210000
[7] 35.2947000 24.7062900 17.2944030 12.1060821 8.4742575 5.9319802
[13] 4.1523862 2.9066703 2.0346692 1.4242685 0.9969879 0.6978915
[19] 0.4885241 0.3419669 0.2393768
Warning, dgeom and friends goes from x=0, not x=1; while you can shift the inputs and outputs to the R functions, it's much easier if you subtract 1 from all your geometric values and test that. I will proceed as if your sample has had 1 subtracted so that it goes from 0.
I'll cut that off at the 15th term (x=14), and group 15+ into its own group (a single group in this case). If you wanted to follow the 'greater than five' rule of thumb, you'd cut it off after the 12th term (x=11). In some cases (such as smaller p), you might want to split the tail across several bins rather than one.
> expec <- dgeom(0:14,.3)*1000
> expec <- c(expec, 1000-sum(expec))
> expec
[1] 300.000000 210.000000 147.000000 102.900000 72.030000 50.421000
[7] 35.294700 24.706290 17.294403 12.106082 8.474257 5.931980
[13] 4.152386 2.906670 2.034669 4.747562
The last cell is the "15+" category. We also need the probabilities.
Now we don't yet have a sample; I'll just generate one:
y <- rgeom(1000,0.3)
but now we want a table of observed counts:
(x <- table(factor(y,levels=0:14),exclude=NULL))
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 <NA>
292 203 150 96 79 59 47 25 16 10 6 7 0 2 5 3
Now you could compute the chi-square directly and then calculate the p-value:
> (chisqstat <- sum((x-expec)^2/expec))
[1] 17.76835
(pval <- pchisq(chisqstat,15,lower.tail=FALSE))
[1] 0.2750401
but you can also get R to do it:
> chisq.test(x,p=expec/1000)
Chi-squared test for given probabilities
data: x
X-squared = 17.7683, df = 15, p-value = 0.275
Warning message:
In chisq.test(x, p = expec/1000) :
Chi-squared approximation may be incorrect
Now the case for unspecified p is similar, but (to my knowledge) you can no longer get chisq.test to do it directly, you have to do it the first way, but you have to estimate the parameter from the data (by maximum likelihood or minimum chi-square), and then test as above but you have one fewer degree of freedom for estimating the parameter.
See the example of doing a chi-square for a Poisson with estimated parameter here; the geometric follows the much same approach as above, with the adjustments as at the link (dealing with the unknown parameter, including the loss of 1 degree of freedom).
Let us assume you've got your randomly-generated variates in a vector x. You can do the following:
x <- rgeom(1000,0.2)
x_tbl <- table(x)
x_val <- as.numeric(names(x_tbl))
x_df <- data.frame(count=as.numeric(x_tbl), value=x_val)
# Expand to fill in "gaps" in the values caused by 0 counts
all_x_val <- data.frame(value = 0:max(x_val))
x_df <- merge(all_x_val, x_df, by="value", all.x=TRUE)
x_df$count[is.na(x_df$count)] <- 0
# Get theoretical probabilities
x_df$eprob <- dgeom(x_df$val, 0.2)
# Chi-square test: once with asymptotic dist'n,
# once with bootstrap evaluation of chi-sq test statistic
chisq.test(x=x_df$count, p=x_df$eprob, rescale.p=TRUE)
chisq.test(x=x_df$count, p=x_df$eprob, rescale.p=TRUE,
simulate.p.value=TRUE, B=10000)
There's a "goodfit" function described as "Goodness-of-fit Tests for Discrete Data" in package "vcd".
G.fit <- goodfit(x, type = "nbinomial", par = list(size = 1))
I was going to use the code you had posted in an earlier question, but it now appears that you have deleted that code. I find that offensive. Are you using this forum to gather homework answers and then defacing it to remove the evidence? (Deleted questions can still be seen by those of us with sufficient rep, and the interface prevents deletion of question with upvoted answers so you should not be able to delete this one.)
Generate a QQ Plot for testing a geometrically distributed sample
--- question---
I have a sample of n elements generated in R with
sim.geometric <- function(nvals)
{
p <- 0.3
u <- runif(nvals)
ceiling(log(u)/log(1-p))
}
for which i want to test its distribution, specifically if it indeed follows a geometric distribution. I want to generate a QQ PLot but have no idea how to.
--------reposted answer----------
A QQ-plot should be a straight line when compared to a "true" sample drawn from a geometric distribution with the same probability parameter. One gives two vectors to the functions which essentially compares their inverse ECDF's at each quantile. (Your attempt is not particularly successful:)
sim.res <- sim.geometric(100)
sim.rgeom <- rgeom(100, 0.3)
qqplot(sim.res, sim.rgeom)
Here I follow the lead of the authors of qqplot's help page (which results in flipping that upper curve around the line of identity):
png("QQ.png")
qqplot(qgeom(ppoints(100),prob=0.3), sim.res,
main = expression("Q-Q plot for" ~~ {G}[n == 100]))
dev.off()
---image not included---
You can add a "line of good fit" by plotting a line through through the 25th and 75th percentile points for each distribution. (I added a jittering feature to this to get a better idea where the "probability mass" was located:)
sim.res <- sim.geometric(500)
qqplot(jitter(qgeom(ppoints(500),prob=0.3)), jitter(sim.res),
main = expression("Q-Q plot for" ~~ {G}[n == 100]), ylim=c(0,max( qgeom(ppoints(500),prob=0.3),sim.res )),
xlim=c(0,max( qgeom(ppoints(500),prob=0.3),sim.res )))
qqline(sim.res, distribution = function(p) qgeom(p, 0.3),
prob = c(0.25, 0.75), col = "red")
I would like to generate correlated variables specified by a correlation matrix.
First I generate the correlation matrix:
require(psych)
require(Matrix)
cor.table <- matrix( sample( c(0.9,-0.9) , 2500 , prob = c( 0.8 , 0.2 ) , repl = TRUE ) , 50 , 50 )
k=1
while (k<=length(cor.table[1,])){
cor.table[1,k]<-0.55
k=k+1
}
k=1
while (k<=length(cor.table[,1])){
cor.table[k,1]<-0.55
k=k+1
}
ind<-lower.tri(cor.table)
cor.table[ind]<-t(cor.table)[ind]
diag(cor.table) <- 1
This correlation matrix is not consistent, therefore, eigenvalue decomposition is impossible.
TO make it consistent I use nearPD:
c<-nearPD(cor.table)
Once this is done I generate the correlated variables:
fit<-principal(c, nfactors=50,rotate="none")
fit$loadings
loadings<-matrix(fit$loadings[1:50, 1:50],nrow=50,ncol=50,byrow=F)
loadings
cases <- t(replicate(50, rnorm(10)) )
multivar <- loadings %*% cases
T_multivar <- t(multivar)
var<-as.data.frame(T_multivar)
cor(var)
However the resulting correlations are far from anything that I specified initially.
Is it not possible to create such correlations or am I doing something wrong?
UPDATE from Greg Snow's comment it became clear that the problem is that my initial correlation matrix is unreasonable.
The question then is how can I make the matrix reasonable. The goal is:
each of the 49 variables should correlate >.5 with the first variable.
~40 of the variables should have a high >.8 correlation with each other
the remaining ~9 variables should have a low or negative correlation with each other.
Is this whole requirement impossible ?
Try using the mvrnorm function from the MASS package rather than trying to construct the variables yourself.
**Edit
Here is a matrix that is positive definite (so it works as a correlation matrix) and comes close to your criteria, you can tweak the values from there (all the Eigen values need to be positive, so you can see how changing a number affects things):
cor.mat <- matrix(0.2,nrow=50, ncol=50)
cor.mat[1,] <- cor.mat[,1] <- 0.55
cor.mat[2:41,2:41] <- 0.9
cor.mat[42:50, 42:50] <- 0.25
diag(cor.mat) <- 1
eigen(cor.mat)$values
Some numerical experimentation based on your specifications above suggests that the generated matrix will never (what never? well, hardly ever ...) be positive definite, but it also doesn't look far from PD with these values (making lcor below negative will almost certainly make things worse ...)
rmat <- function(n=49,nhcor=40,hcor=0.8,lcor=0) {
m <- matrix(lcor,n,n) ## fill matrix with 'lcor'
## select high-cor variables
hcorpos <- sample(n,size=nhcor,replace=FALSE)
## make all of these highly correlated
m[hcorpos,hcorpos] <- hcor
## compute min real part of eigenvalues
min(Re(eigen(m,only.values=TRUE)$values))
}
set.seed(101)
r <- replicate(1000,rmat())
## NEVER pos definite
max(r)
## [1] -1.069413e-15
par(las=1,bty="l")
png("eighist.png")
hist(log10(abs(r)),breaks=50,col="gray",main="")
dev.off()
I'm looking for some algorithm such as k-means for grouping points on a map into a fixed number of groups, by distance.
The number of groups has already been decided, but the trick part (at least for me) is to meet the criteria that the sum of MOS of each group should in the certain range, say bigger than 1. Is there any way to make that happen?
ID MOS X Y
1 0.47 39.27846 -76.77101
2 0.43 39.22704 -76.70272
3 1.48 39.24719 -76.68485
4 0.15 39.25172 -76.69729
5 0.09 39.24341 -76.69884
I was intrigued by your question but was unsure how you might introduce some sort of random process into a grouping algorithm. Seems that the kmeans algorithm does indeed give different results if you permutate your dataset (e.g. the order of the rows). I found this bit of info here. The following script demonstrates this with a random set of data. The plot shows the raw data in black and then draws a segment to the center of each cluster by permutation (colors).
Since I'm not sure how your MOS variable is defined, I have added a random variable to the dataframe to illustrate how you might look for clusterings that satisfy a given criteria. The sum of MOS is calculated for each cluster and the result is stored in the MOS.sums object. In order to reproduce a favorable clustering, you can use the random seed value that was used for the permutation, which is stored in the seeds object. You can see that the permutations result is several different clusterings:
set.seed(33)
nsamples=500
nperms=10
nclusters=3
df <- data.frame(x=runif(nsamples), y=runif(nsamples), MOS=runif(nsamples))
MOS.sums <- matrix(NaN, nrow=nperms, ncol=nclusters)
colnames(MOS.sums) <- paste("cluster", 1:nclusters, sep=".")
rownames(MOS.sums) <- paste("perm", 1:nperms, sep=".")
seeds <- round(runif(nperms, min=1, max=10000))
plot(df$x, df$y)
COL <- rainbow(nperms)
for(i in seq(nperms)){
set.seed(seeds[i])
ORD <- sample(nsamples)
K <- kmeans(df[ORD,1:2], centers=nclusters)
MOS.sums[i,] <- tapply(df$MOS[ORD], K$cluster, sum)
segments(df$x[ORD], df$y[ORD], K$centers[K$cluster,1], K$centers[K$cluster,2], col=COL[i])
}
seeds
MOS.sums
In the following code I use bootstrapping to calculate the C.I. and the p-value under the null hypothesis that two different fertilizers applied to tomato plants have no effect in plants yields (and the alternative being that the "improved" fertilizer is better). The first random sample (x) comes from plants where a standard fertilizer has been used, while an "improved" one has been used in the plants where the second sample (y) comes from.
x <- c(11.4,25.3,29.9,16.5,21.1)
y <- c(23.7,26.6,28.5,14.2,17.9,24.3)
total <- c(x,y)
library(boot)
diff <- function(x,i) mean(x[i[6:11]]) - mean(x[i[1:5]])
b <- boot(total, diff, R = 10000)
ci <- boot.ci(b)
p.value <- sum(b$t>=b$t0)/b$R
What I don't like about the code above is that resampling is done as if there was only one sample of 11 values (separating the first 5 as belonging to sample x leaving the rest to sample y).
Could you show me how this code should be modified in order to draw resamples of size 5 with replacement from the first sample and separate resamples of size 6 from the second sample, so that bootstrap resampling would mimic the “separate samples” design that produced the original data?
EDIT2 :
Hack deleted as it was a wrong solution. Instead one has to use the argument strata of the boot function :
total <- c(x,y)
id <- as.factor(c(rep("x",length(x)),rep("y",length(y))))
b <- boot(total, diff, strata=id, R = 10000)
...
Be aware you're not going to get even close to a correct estimate of your p.value :
x <- c(1.4,2.3,2.9,1.5,1.1)
y <- c(23.7,26.6,28.5,14.2,17.9,24.3)
total <- c(x,y)
b <- boot(total, diff, strata=id, R = 10000)
ci <- boot.ci(b)
p.value <- sum(b$t>=b$t0)/b$R
> p.value
[1] 0.5162
How would you explain a p-value of 0.51 for two samples where all values of the second are higher than the highest value of the first?
The above code is fine to get a -biased- estimate of the confidence interval, but the significance testing about the difference should be done by permutation over the complete dataset.
Following John, I think the appropriate way to use bootstrap to test if the sums of these two different populations are significantly different is as follows:
x <- c(1.4,2.3,2.9,1.5,1.1)
y <- c(23.7,26.6,28.5,14.2,17.9,24.3)
b_x <- boot(x, sum, R = 10000)
b_y <- boot(y, sum, R = 10000)
z<-(b_x$t0-b_y$t0)/sqrt(var(b_x$t[,1])+var(b_y$t[,1]))
pnorm(z)
So we can clearly reject the null that they are the same population. I may have missed a degree of freedom adjustment, I am not sure how bootstrapping works in that regard, but such an adjustment will not change your results drastically.
While the actual soil beds could be considered a stratified variable in some instances this is not one of them. You only have the one manipulation, between the groups of plants. Therefore, your null hypothesis is that they really do come from the exact same population. Treating the items as if they're from a single set of 11 samples is the correct way to bootstrap in this case.
If you have two plots, and in each plot tried the different fertilizers over different seasons in a counterbalanced fashion then the plots would be statified samples and you'd want to treat them as such. But that isn't the case here.