Peterson_-_Machine_Learning_Project

Machine Learning Project
Joshua Peterson
10/4/2016
Background
In this project I analyzed data from accelerometers on the belt, forearm, arm, and dumbell of 6 participants.
This data may have come from IoT, fitness devices such as Jawbone Up, Nike FuelBand and the Fitbit. Each
participant was asked to perform barbell lifts correctly and incorrectly in 5 different ways.
The intent of the project is to predict the manner in which each of the participants performed each of the
performed excercies. This variable is the “classe” variable in the training set.
The training data for the project is available at:
https://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv
The test data for the project is availabe at:
https://d396qusza40orc.cloudfront.net/predmachlearn/pml-testing.csv
Study Design
I will be following the standard Prediction Design Framework
1. Define error rate
2. Splitting the data in to Training, Testing and Validation
3. Choose features of the Training data set using cross-validation
4. Choose prediction function of the Training data using cross-validation
5. If there is no validation apply 1x to test set
6. If there is validation apply to test set and refine and apply 1x to validation
It appears that we have a relatively large sample size therefore I would like to target the following parameters:
1. 60% training
2. 20% test
3. 20% validation
Loading the excercise data.
Data Splitting
In this data splitting action, I am splitting the data into 60% training set and 40% testing set.
library(caret); data(train_data)
## Error in read.table(file = file, header = header, sep = sep, quote = quote, : 'file' must be a charac
inTrain <- createDataPartition(y=train_data$classe,p=0.60, list=FALSE)
training <- train_data[inTrain,]
testing <- train_data[-inTrain,]
dim(training); dim(testing)
## [1] 11776 160
1

## [1] 7846 160
Cleaning the data
Because there are some non-zero values within the data set, we must clean the data set to extract those
values.
train_data_NZV<- nearZeroVar(training, saveMetrics = TRUE)
training<- training[,train_data_NZV$nzv==FALSE]
test_data_NZV<- nearZeroVar(testing, saveMetrics = TRUE)
testing<- testing[, test_data_NZV$nzv==FALSE]
Next, because there are data points within the ﬁrst column that may interfere with my alogorithms, I will
remove the ﬁrst column from the data set
training<- training[c(-1)]
Next, I will elimiate those variables with an excessive amount of NAs values.
final_training <- training
for(i in 1:length(training)) {
if( sum( is.na( training[, i] ) ) /nrow(training) >= .6 ) {
for(j in 1:length(final_training)) {
if( length( grep(names(training[i]), names(final_training)[j]) ) ==1) { fin
}
}
}
}
dim(final_training)
## [1] 11776 58
training<- final_training
rm(final_training)
I will now perform the same data cleansing process for the testing data that I performed for the training data.
clean_training<- colnames(training)
clean_training_2<- colnames(training[, -58])
testing<- testing[clean_training]
testing_2<- testing_2[clean_training_2]
## Error in eval(expr, envir, enclos): object 'testing_2' not found
dim(testing)
## [1] 7846 58
Lastly, to ensure proper function of the Decision Tree analysis, we must coerce the data
for (i in 1:length(testing) ) {
for(j in 1:length(training)) {
if( length( grep(names(training[i]), names(testing)[j]) ) ==1) {
class(testing[j]) <- class(training[i])
}
}
2

}
testing <- rbind(training[2, -58] , testing)
## Error in rbind(deparse.level, ...): numbers of columns of arguments do not match
testing <- testing[-1,]
K-Fold Cross Validation
We will then use the K-Fold process to cross-validate the data by splitting the training set in to many, smaller
data sets.
Here, I am creating 10 folds and setting a random number seed of 32323 for the study. Each fold has
approximately the same number of samples in it.
set.seed(32323)
folds <- createFolds(y=train_data$classe,k=10,list=TRUE,returnTrain=TRUE)
sapply(folds,length)
## Fold01 Fold02 Fold03 Fold04 Fold05 Fold06 Fold07 Fold08 Fold09 Fold10
## 17660 17660 17661 17660 17659 17658 17660 17660 17660 17660
folds[[1]][1:10]
## [1] 1 2 3 4 5 6 7 8 9 10
Here, I wanted to resample the data set
set.seed(32323)
folds <- createResample(y=train_data$classe,times=10,
list=TRUE)
sapply(folds,length)
## Resample01 Resample02 Resample03 Resample04 Resample05 Resample06
## 19622 19622 19622 19622 19622 19622
## Resample07 Resample08 Resample09 Resample10
## 19622 19622 19622 19622
folds[[1]][1:10]
## [1] 2 3 3 4 6 6 7 8 9 10
Machine learning alogorithm decisioning
First I wanted to determine the optimal machine learning model to use. The first test I used the Decision
Tree approach. I followed up by testing the Random Forest approach.
Machine learning using Decision Trees
The first task was to determine model fit.
Next, construct a Decision Tree graph
fancyRpartPlot(modelFit)
Next, I used the predict function for model fitting.
3

Figure 1: plot of chunk unnamed-chunk-10
4

predict_perf<- predict(modelFit, testing, type = "class")
Lastly, I used a confusion matrix to test the results.
confusionMatrix(predict_perf, testing$classe)
## Confusion Matrix and Statistics
##
## Reference
## Prediction A B C D E
## A 2154 62 3 3 0
## B 58 1237 67 60 0
## C 19 206 1271 144 58
## D 0 13 27 1011 166
## E 0 0 0 68 1218
##
## Overall Statistics
##
## Accuracy : 0.8784
## 95% CI : (0.871, 0.8855)
## No Information Rate : 0.2844
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.8463
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: A Class: B Class: C Class: D Class: E
## Sensitivity 0.9655 0.8149 0.9291 0.7862 0.8447
## Specificity 0.9879 0.9708 0.9341 0.9686 0.9894
## Pos Pred Value 0.9694 0.8699 0.7485 0.8307 0.9471
## Neg Pred Value 0.9863 0.9563 0.9842 0.9585 0.9658
## Prevalence 0.2844 0.1935 0.1744 0.1639 0.1838
## Detection Rate 0.2746 0.1577 0.1620 0.1289 0.1553
## Detection Prevalence 0.2832 0.1813 0.2164 0.1551 0.1639
## Balanced Accuracy 0.9767 0.8928 0.9316 0.8774 0.9170
The output for this approach is not too bad. Overall model accuracy was 0.8711 or 87.11%. Within a 95%
probability, the model accuracy ranges between 0.8635 and 0.8785. This test is conﬁrmed by a p-value < 0.05.
Machine learning using Random Forests
Like the previous analysis, I wanted to determine model ﬁt. This time instead of using the Decision Tree
approach, I used the randomForest function.
modelFit2<- randomForest(classe ~., data = training)
Next, I wanted to predict the in-sample error.
predict_perf2<- predict(modelFit2, testing, type = "class")
And once again, the last step was to use a confusion matrix to test results.
confusionMatrix(predict_perf2, testing$classe)
5

##
## Reference
## A 2231 0 0 0 0
## B 0 1518 1 0 0
## C 0 0 1367 3 0
## D 0 0 0 1279 0
## E 0 0 0 4 1442
##
##
## Accuracy : 0.999
## 95% CI : (0.998, 0.9996)
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.9987
##
##
## Sensitivity 1.0000 1.0000 0.9993 0.9946 1.0000
## Specificity 1.0000 0.9998 0.9995 1.0000 0.9994
## Pos Pred Value 1.0000 0.9993 0.9978 1.0000 0.9972
## Neg Pred Value 1.0000 1.0000 0.9998 0.9989 1.0000
## Prevalence 0.2844 0.1935 0.1744 0.1639 0.1838
## Detection Rate 0.2844 0.1935 0.1743 0.1630 0.1838
Model accuracy is 0.9994 or 99.94% with a 95% probability that the model has an accuracy between 0.9985
and 0.9998. This test is conﬁrmed by a p-value of < 0.05.
plot(modelFit2)
Machine learning using Boosted Regressions
set.seed(2058)
fitControl <- trainControl(method = "repeatedcv",number = 5, repeats = 1)
gbmFit <- train(classe ~ ., data=training, method = "gbm",
trControl = fitControl,
verbose = FALSE)
gbmFinMod <- gbmFit$finalModel
gbmPredTest <- predict(gbmFit, newdata=testing)
gbmAccuracyTest <- confusionMatrix(gbmPredTest, testing$classe)
gbmAccuracyTest
6

7

##
## Reference
## A 2231 1 0 0 0
## B 0 1513 0 0 0
## C 0 3 1354 6 0
## D 0 1 14 1275 1
## E 0 0 0 5 1441
##
##
## Accuracy : 0.996
## 95% CI : (0.9944, 0.9973)
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.995
##
##
## Sensitivity 1.0000 0.9967 0.9898 0.9914 0.9993
## Specificity 0.9998 1.0000 0.9986 0.9976 0.9992
## Pos Pred Value 0.9996 1.0000 0.9934 0.9876 0.9965
## Neg Pred Value 1.0000 0.9992 0.9978 0.9983 0.9998
## Prevalence 0.2844 0.1935 0.1744 0.1639 0.1838
## Detection Rate 0.2844 0.1929 0.1726 0.1625 0.1837
Model accuracy is 0.996 or 99.6% with a 95% probability that model has an accuracy between 0.9944 and
0.9973.
plot(gbmFit, ylim=c(0.80, 1))
After running the Decision Tree, Random Forest and GBM frameworks, I’ve come to the conclusion that
Random Forests is the optimal approach.
Decision Tree error rate: 12.89% Random Forests error rate: 0.06% General Boosted Regressions error rate:
0.40%
Finally, we will use the Random Forest model for prediction
predict_final<- predict(modelFit2, testing, type = "class")
##
##
## processing file: Peterson - Machine Learning Project.Rmd
## Error in parse_block(g[-1], g[1], params.src): duplicate label 'setup'
8

9

Peterson_-_Machine_Learning_Project

More Related Content

Similar to Peterson_-_Machine_Learning_Project (20)

Peterson_-_Machine_Learning_Project