I find different numeric values for the computation of the Area Under the Precision Recall Curve (PRAUC) with the dataset I am working on when computed via 2 different R packages: yardstick and caret.
I am afraid I was not able to reproduce this mismatch with synthetic data, but only with my dataset (this is strange as well)
In order to make this reproducible, I am sharing the prediction output of my model, you can download it here https://drive.google.com/open?id=1LuCcEw-RNRcdz6cg0X5bIEblatxH4Rdz (don't worry, it's a small csv).
The csv contains a dataframe with 4 columns:
yes probability estimate of being in class yes
no = 1 - yes
obs actual class label
pred predicted class label (with .5 threshold)
here follows the code to produce the 2 values of PRAUC
require(data.table)
require(yardstick)
require(caret)
pr <- fread('pred_sample.csv')
# transform to factors
# put the positive class in the first level
pr[, obs := factor(obs, levels = c('yes', 'no'))]
pr[, pred := factor(pred, levels = c('yes', 'no'))] # this is actually not needed
# compute yardstick PRAUC
pr_auc(pr, obs, yes) # 0.315
# compute caret PRAUC
prSummary(pr, lev = c('yes', 'no')) # 0.2373
I could understand a little difference, due to the approximation when computing the area (interpolating the curve), but this seems way too high.
I even tried a third package, PRROC, and the result is still different, namely around .26.
Related
I have a data set with around 130000 records. The records divided in two class of target variable,0 & 1. 1 contains only 0.09% of total proportion.
I'm running my analysis in R-3.5.1 on Windows 10. I used SMOTE algorithm to work with this imbalanced data set.
I used following code to handle imbalanced data set
library(DMwR)
data_code$target=as.factor(data_code$target) #Converted to factor as
# SMOTE works with factor data type
smoted_data <- SMOTE(target~., data_code, perc.over=100)
But after executing the code,I'm seeing the count for 0 is 212 & 1 is also 212 which is significant reduction of my sample size.Can you suggest me how do I handle this imbalanced data set with SMOTE without changing my data size
You need to play a bit with the two parameters avaiable from the function: perc.over and perc.under.
As per the doc from SMOTE:
The parameters perc.over and perc.under control the amount of
over-sampling of the minority class and under-sampling of the majority
classes, respectively.
So:
perc.over will tipically be a number above 100. With this type of
values, for each case in the orginal data set belonging to the
minority class, perc.over/100 new examples of that class will be
created
I can't see your data but, if your minority class has 100 cases and perc.over=100, the algorithm will generate 100/100 = 1 new cases from that class.
The parameter perc.under controls the proportion of cases of the
majority class that will be randomly selected for the final "balanced"
data set. This proportion is calculated with respect to the number of
newly generated minority class cases.
So for example a value of perc.under=100 will select from the majority class on the original data the same amount of observation that have been generated for the minority class.
In our example just 1 new case was generated so it will add just another one, resulting in a new dataset with 2 cases.
I suggest to use values above 100 for perc.over, and an even higher value for perc.under (defaults are 100 and 200).
Keep in mind that you're adding new observations that are not real in your minority class, I'd try to keep these under control.
Numeric example:
set.seed(123)
data <- data.frame(var1 = sample(50),
var2 = sample(50),
out = as.factor(rbinom(50, 1, prob=0.1)))
table(data$out)
# 0 1
# 43 7 # 50 rows total (original data)
smote_data <- DMwR::SMOTE(out ~ var1, data, perc.over = 200, perc.under = 400)
table(smote_data$out)
# 0 1
# 56 21 # 77 rows total (smote data)
An alternative to the DMwR package is the smotefamily package which does not reduce the sample size.
Instead, it creates additional data (= synthesized data) from the minority class, and adds it to the original data. So the output in the $data argument is ready for training. To tune the amount of synthesized data, you can modify the parameter dup_size. However, the default dup_size = 0 already optimizes the output to achieve balanced classes, so you don't need to tune it.
This is greatly explained in this blog post by Richard Richard.
Example code (with features in first two columns):
smote1 <- smotefamily::SMOTE(features, target, K = 4, dup_size = 0)
formula1 <- "class ~ ." %>% as.formula
model.smote <- caret::train(formula1, method = "rpart", smote1$data)
predictions.smote <- predict(model.smote, smote1$data[,1:2]) %>% print
cv2 <- confusionMatrix(smote1$data$class %>% as.factor, predictions.smote)
I find the smotefamily::SMOTE more convenient because you don't have to tune the two parameters perc_over and perc_under until you get an acceptable sample size, and the DMwR::SMOTE often generates NA values.
I know I'm a little too late to answer your question but hope this answer would help others! The package you're using is DMwR which uses a combination of SMOTE and under-sampling of the majority class.
I'd suggest you to use smotefamily::SMOTE as it only over samples the minority class, so you wouldn't lose your majority class observations.
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.
I received some good help getting my data formatted properly produce a multinomial logistic model with mlogit here (Formatting data for mlogit)
However, I'm trying now to analyze the effects of covariates in my model. I find the help file in mlogit.effects() to be not very informative. One of the problems is that the model appears to produce a lot of rows of NAs (see below, index(mod1) ).
Can anyone clarify why my data is producing those NAs?
Can anyone help me get mlogit.effects to work with the data below?
I would consider shifting the analysis to multinom(). However, I can't figure out how to format the data to fit the formula for use multinom(). My data is a series of rankings of seven different items (Accessible, Information, Trade offs, Debate, Social and Responsive) Would I just model whatever they picked as their first rank and ignore what they chose in other ranks? I can get that information.
Reproducible code is below:
#Loadpackages
library(RCurl)
library(mlogit)
library(tidyr)
library(dplyr)
#URL where data is stored
dat.url <- 'https://raw.githubusercontent.com/sjkiss/Survey/master/mlogit.out.csv'
#Get data
dat <- read.csv(dat.url)
#Complete cases only as it seems mlogit cannot handle missing values or tied data which in this case you might get because of median imputation
dat <- dat[complete.cases(dat),]
#Change the choice index variable (X) to have no interruptions, as a result of removing some incomplete cases
dat$X <- seq(1,nrow(dat),1)
#Tidy data to get it into long format
dat.out <- dat %>%
gather(Open, Rank, -c(1,9:12)) %>%
arrange(X, Open, Rank)
#Create mlogit object
mlogit.out <- mlogit.data(dat.out, shape='long',alt.var='Open',choice='Rank', ranked=TRUE,chid.var='X')
#Fit Model
mod1 <- mlogit(Rank~1|gender+age+economic+Job,data=mlogit.out)
Here is my attempt to set up a data frame similar to the one portrayed in the help file. It doesnt work. I confess although I know the apply family pretty well, tapply is murky to me.
with(mlogit.out, data.frame(economic=tapply(economic, index(mod1)$alt, mean)))
Compare from the help:
data("Fishing", package = "mlogit")
Fish <- mlogit.data(Fishing, varying = c(2:9), shape = "wide", choice = "mode")
m <- mlogit(mode ~ price | income | catch, data = Fish)
# compute a data.frame containing the mean value of the covariates in
# the sample data in the help file for effects
z <- with(Fish, data.frame(price = tapply(price, index(m)$alt, mean),
catch = tapply(catch, index(m)$alt, mean),
income = mean(income)))
# compute the marginal effects (the second one is an elasticity
effects(m, covariate = "income", data = z)
I'll try Option 3 and switch to multinom(). This code will model the log-odds of ranking an item as 1st, compared to a reference item (e.g., "Debate" in the code below). With K = 7 items, if we call the reference item ItemK, then we're modeling
log[ Pr(Itemk is 1st) / Pr(ItemK is 1st) ] = αk + xTβk
for k = 1,...,K-1, where Itemk is one of the other (i.e. non-reference) items. The choice of reference level will affect the coefficients and their interpretation, but it will not affect the predicted probabilities. (Same story for reference levels for the categorical predictor variables.)
I'll also mention that I'm handling missing data a bit differently here than in your original code. Since my model only needs to know which item gets ranked 1st, I only need to throw out records where that info is missing. (E.g., in the original dataset record #43 has "Information" ranked 1st, so we can use this record even though 3 other items are NA.)
# Get data
dat.url <- 'https://raw.githubusercontent.com/sjkiss/Survey/master/mlogit.out.csv'
dat <- read.csv(dat.url)
# dataframe showing which item is ranked #1
ranks <- (dat[,2:8] == 1)
# for each combination of predictor variable values, count
# how many times each item was ranked #1
dat2 <- aggregate(ranks, by=dat[,9:12], sum, na.rm=TRUE)
# remove cases that didn't rank anything as #1 (due to NAs in original data)
dat3 <- dat2[rowSums(dat2[,5:11])>0,]
# (optional) set the reference levels for the categorical predictors
dat3$gender <- relevel(dat3$gender, ref="Female")
dat3$Job <- relevel(dat3$Job, ref="Government backbencher")
# response matrix in format needed for multinom()
response <- as.matrix(dat3[,5:11])
# (optional) set the reference level for the response by changing
# the column order
ref <- "Debate"
ref.index <- match(ref, colnames(response))
response <- response[,c(ref.index,(1:ncol(response))[-ref.index])]
# fit model (note that age & economic are continuous, while gender &
# Job are categorical)
library(nnet)
fit1 <- multinom(response ~ economic + gender + age + Job, data=dat3)
# print some results
summary(fit1)
coef(fit1)
cbind(dat3[,1:4], round(fitted(fit1),3)) # predicted probabilities
I didn't do any diagnostics, so I make no claim that the model used here provides a good fit.
You are working with Ranked Data, not just Multinomial Choice Data. The structure for the Ranked data in mlogit is that first set of records for a person are all options, then the second is all options except the one ranked first, and so on. But the index assumes equal number of options each time. So a bunch of NAs. We just need to get rid of them.
> with(mlogit.out, data.frame(economic=tapply(economic, index(mod1)$alt[complete.cases(index(mod1)$alt)], mean)))
economic
Accessible 5.13
Debate 4.97
Information 5.08
Officials 4.92
Responsive 5.09
Social 4.91
Trade.Offs 4.91
I want to measure the features importance with the cforest function from the party library.
My output variable has something like 2000 samples in class 0 and 100 samples in class 1.
I think a good way to avoid bias due to class unbalance is to train each tree of the forest using a subsample such that the number of elements of class 1 is the same of the number of element in class 0.
Is there anyway to do that? I am thinking to an option like n_samples = c(20, 20)
EDIT:
An example of code
> iris.cf <- cforest(Species ~ ., data = iris,
+ control = cforest_unbiased(mtry = 2)) #<--- Here I would like to train the forest using a balanced subsample of the data
> varimp(object = iris.cf)
Sepal.Length Sepal.Width Petal.Length Petal.Width
0.048981818 0.002254545 0.305818182 0.271163636
>
EDIT:
Maybe my question is not clear enough.
Random forest is a set of decision trees. In general the decision trees are constructed using only a random subsample of the data. I would like that the used subsample has the same numbers of element in the class 1 and in the class 0.
EDIT:
The function that I am looking for is for sure available in the randomForest package
sampsize
Size(s) of sample to draw. For classification, if sampsize is a vector of the length the number of strata, then sampling is stratified by strata, and the elements of sampsize indicate the numbers to be drawn from the strata.
I need the same for the party package. Is there any way to get it?
I will assume you know what you want to accomplish, but don't know enough R to do that.
Not sure if the function provides balancing of data as an argument, but you can do it manually. Below is the code I quickly threw together. More elegant solution might exist.
# just in case
myData <- iris
# replicate everything *10* times. Replicate is just a "loop 10 times".
replicate(10,
{
# split dataset by class and add separate classes to list
splitList <- split(myData, myData$Species)
# sample *20* random rows from each matrix in a list
sampledList <- lapply(splitList, function(dat) { dat[sample(20),] })
# combine sampled rows to a data.frame
sampledData <- do.call(rbind, sampledList)
# your code below
res.cf <- cforest(Species ~ ., data = sampledData,
control = cforest_unbiased(mtry = 2)
)
varimp(object = res.cf)
}
)
Hope you can take it from here.
In the following code I use bootstrapping to calculate the C.I. and the p-value under the null hypothesis that two different fertilizers applied to tomato plants have no effect in plants yields (and the alternative being that the "improved" fertilizer is better). The first random sample (x) comes from plants where a standard fertilizer has been used, while an "improved" one has been used in the plants where the second sample (y) comes from.
x <- c(11.4,25.3,29.9,16.5,21.1)
y <- c(23.7,26.6,28.5,14.2,17.9,24.3)
total <- c(x,y)
library(boot)
diff <- function(x,i) mean(x[i[6:11]]) - mean(x[i[1:5]])
b <- boot(total, diff, R = 10000)
ci <- boot.ci(b)
p.value <- sum(b$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.