I used a subset of data (https://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv) from the http://groupware.les.inf.puc-rio.br/har#dataset#ixzz4cSvunXAe to predict classe, which range from A to E and correspond to various ways of performing bicep curls.
train1<-read.csv("C:/Users/Owner/Documents/pml-training.csv")
library(caret)
The starting dataset has 19622 observations and 160 variables.
dim(train1)
## [1] 19622 160
I first reduced the number of columns in the dataset by removing columns with near zero variances. I took out all the columns with NA values because each of them had 19,216 NA values. Due to each of these previously mentioned columns only having numerical values for two percent of the observations, I decided to omit them as predictors.
nearzero<-nearZeroVar(train1)
train2<-train1[,-nearzero]
findna<-colSums(sapply(train2, is.na))
head(findna[findna>0])
## max_roll_belt max_picth_belt min_roll_belt
## 19216 19216 19216
## min_pitch_belt amplitude_roll_belt amplitude_pitch_belt
## 19216 19216 19216
train2<-train2[,findna==0]
I also got rid of the first five columns which dealt with observation number (X), username, and timestamps as well as num_window (the sixth column). As shown in the plot below, num_window increases similarly as observation number goes up within each classe.
qplot(X, num_window, data = train2, color=classe)
train2<-train2[,-c(1:6)]
Now 53 variables remain.
I split 75 percent of the observations into a training set and the rest into a validation set.
set.seed(325)
inTrain<-createDataPartition(train2$classe, p=0.75, list=FALSE)
traindata<-train2[inTrain,]
validdata<-train2[-inTrain,]
A random forest algorithm was used to predict classe using the remaining 52 variables because it is one of the most accurate algorithm. However, I used cross validation with 10 subsamples within the trainControl attribute of the train function to avoid overfitting.
set.seed(325)
modelFit<-train(classe~., data=traindata, method="rf", trControl= trainControl(method="cv", number=10, allowParallel = TRUE))
The model was used to predict the classe of the observations in the validation set. The Accuracy of the model on the validation data is 99.55%, therefore I estimate out-of-sample error to be a low 0.45%.
predmod<-predict(modelFit, validdata)
confusionMatrix(predmod, validdata$classe)
## Confusion Matrix and Statistics
##
## Reference
## Prediction A B C D E
## A 1395 3 0 0 0
## B 0 945 7 0 0
## C 0 1 848 9 0
## D 0 0 0 794 1
## E 0 0 0 1 900
##
## Overall Statistics
##
## Accuracy : 0.9955
## 95% CI : (0.9932, 0.9972)
## No Information Rate : 0.2845
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.9943
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: A Class: B Class: C Class: D Class: E
## Sensitivity 1.0000 0.9958 0.9918 0.9876 0.9989
## Specificity 0.9991 0.9982 0.9975 0.9998 0.9998
## Pos Pred Value 0.9979 0.9926 0.9883 0.9987 0.9989
## Neg Pred Value 1.0000 0.9990 0.9983 0.9976 0.9998
## Prevalence 0.2845 0.1935 0.1743 0.1639 0.1837
## Detection Rate 0.2845 0.1927 0.1729 0.1619 0.1835
## Detection Prevalence 0.2851 0.1941 0.1750 0.1621 0.1837
## Balanced Accuracy 0.9996 0.9970 0.9947 0.9937 0.9993
Data source: Ugulino, W.; Cardador, D.; Vega, K.; Velloso, E.; Milidiu, R.; Fuks, H. Wearable Computing: Accelerometers' Data Classification of Body Postures and Movements. Proceedings of 21st Brazilian Symposium on Artificial Intelligence. Advances in Artificial Intelligence - SBIA 2012. In: Lecture Notes in Computer Science. , pp. 52-61. Curitiba, PR: Springer Berlin / Heidelberg, 2012. ISBN 978-3-642-34458-9. DOI: 10.1007/978-3-642-34459-6_6.