The Human Activity Recognition (HAR) project collected exercise data, focusing on the quality of exercise related to specific motion types. A more detailed explanation of the data collection process and project overview can be found at the HAR website.

We are trying to predict the “classe” variable, which assigns one of five letter grades to the exercise - “A” (the best), B, C, D and “E” (the worst). This report outlines the process in building the model we use to predict “classe”.


Establish a Project Folder

We begin by establishing a project folder (if one hasn’t been established already): 

mainDir <- "C:\\Users\\Bob\\Documents\\R\\08 Practical Machine Learning"
subDir <- "Exercise Project"

if(file.exists(subDir)){
  setwd(file.path(mainDir, subDir))
} else {
  dir.create(file.path(mainDir, subDir), showWarnings = FALSE)
  setwd(file.path(mainDir, subDir))
}


Load Relevant Packages

The following packages will be used: 

library(caret)
library(corrplot)


Download and Read Data

The training and test datasets - each have 160 variables with 19,622 and 20 observations, respectively - are downloaded and read into R as follows: 

#DOWNLOAD & READ TRAINING DATA:
download.file("https://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv",
              destfile="trainData")
trainOrig <- read.csv("./trainData", header = TRUE, na.strings = c("NA", ""))

#DOWNLOAD & READ TEST DATA
download.file("https://d396qusza40orc.cloudfront.net/predmachlearn/pml-testing.csv",
              destfile="testData")
testOrig <- read.csv("./testData", header = TRUE, na.strings = c("NA", ""))


Note that in loading the object, blanks are converted to “NA”. This needed to execute a subsequent step.


Cleaning Data

The following steps convert a raw dataset into a workable one.

First, all “NA” values - whether native to the original dataset or coerced - are removed. The object cSum is a numeric object that summarizes the number of “NAs” in each column: 

cSum <- colSums(is.na(trainOrig))


Next, original variables are subsetted to include variables containing non-NA data (via the argument cSum==0). This same command is applied to both the training and test sets, which have column consistency. This reduces the number of explanatory variables from 160 to 60. 

trainClean <- trainOrig[, cSum == 0]
testClean <- testOrig[, cSum == 0]


Lastly, we further purify the dataset by removing columns that have little-to-no predictive relevance. The following commands reduce the size of the datasets from 60 to 54 variables: 

useless <- grepl("X|user_name|raw_timestamp_part_1|raw_timestamp_part_2|
                 |cvtd_timestamp|new_window",
                 colnames(trainClean))
trainClean <- trainClean[, !useless]
testClean <- testClean[, !useless]


Variable Correlation and Model Selection

Prior to model development, we take an exploratory look at variable correlation. As a general rule, we want to use as few variables with the most explanatory power as possible. The matrix below depicts the relationship between variables: 

corMat <- cor(trainClean[, -54])
corrplot(corMat, order = "FPC", 
         method = "color",
         type = "lower",
         tl.cex = 0.4,
         tl.col = rgb(0,0,0))


The correlation matrix shows that generally, multicollinearity is not a huge issue across the independent variable set. This, along with the categorical nature of the dependant (classe) variable, we’ll approach this problem using a random forest (RF) model. RF models tend to be quite accurate; however, tend to be slower, and in the absence of cross-validation, are prone to overfitting.

Create Validation Dataset

Now that we have a cleaned, workable dataset, we can split-off a validation set from the training set: 

inTrain = createDataPartition(y = trainClean$classe,
                              p = 0.75,
                              list = FALSE)
trainSet = trainClean[inTrain, ]
validSet = trainClean[-inTrain, ]


Model Creation & Validation

We next train the random forest model. In this model, 4-fold cross-validation is used. Due to the computational intensity of this type of model, this step may take a relatively long time to execute: 

ctrlRf <- trainControl(method = "cv", 4)
modelRf <- train(classe ~ ., 
                 data = trainSet,
                 method = "rf",
                 trControl = ctrlRf,
                 ntree = 150)


Now that the model is trained, we test the model on the validation set. Model performance can be assessed by generating a confusion matrix: 

predictRf <- predict(modelRf, validSet)
confMatrix <- confusionMatrix(validSet$classe, predictRf)
confMatrix
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction    A    B    C    D    E
##          A 1395    0    0    0    0
##          B    2  947    0    0    0
##          C    0    5  850    0    0
##          D    0    0    0  803    1
##          E    0    0    0    2  899
## 
## Overall Statistics
##                                          
##                Accuracy : 0.998          
##                  95% CI : (0.9963, 0.999)
##     No Information Rate : 0.2849         
##     P-Value [Acc > NIR] : < 2.2e-16      
##                                          
##                   Kappa : 0.9974         
##  Mcnemar's Test P-Value : NA             
## 
## Statistics by Class:
## 
##                      Class: A Class: B Class: C Class: D Class: E
## Sensitivity            0.9986   0.9947   1.0000   0.9975   0.9989
## Specificity            1.0000   0.9995   0.9988   0.9998   0.9995
## Pos Pred Value         1.0000   0.9979   0.9942   0.9988   0.9978
## Neg Pred Value         0.9994   0.9987   1.0000   0.9995   0.9998
## Prevalence             0.2849   0.1941   0.1733   0.1642   0.1835
## Detection Rate         0.2845   0.1931   0.1733   0.1637   0.1833
## Detection Prevalence   0.2845   0.1935   0.1743   0.1639   0.1837
## Balanced Accuracy      0.9993   0.9971   0.9994   0.9986   0.9992


The confusion matrix above shows the actual values along the horizontal and the predicted values along the vertical. Only 6 false-negative and 2 false-positive errors were made. The accuracy and estimated out-of-sample error are 99.84% and 0.16%, respectively.


plot(confMatrix$table,
     col = 67,
     main = paste("Confusion Matrix: Accuracy = ",
                  round(confMatrix$overall["Accuracy"], 4)))


Prediction on Test Set

To reiterate, the purpose of this project is to predict the Classe variable of the 20-observation test set using the validated model. The following commands achieve this objective. Note the “problem_id” column must first be removed: 

result <- predict(modelRf, testClean[, -54])
result
##  [1] B A B A A E D B A A B C B A E E A B B B
## Levels: A B C D E


The model successfully predicts each of the 20 Classe values.


Citation:
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.