Assignment

For this assignment, we have been given two files - one with training data and one with testing data. The data is taken from a Human Activity Recognition (HAR) study. The goal of this assignment is to predict the classe of each test example, based on the available motion sensor data.

Pre-Processing

knitr::opts_chunk$set(echo = TRUE, results = 'markup',
                      warning = FALSE, message = FALSE,
                      cache = TRUE)

library(ggplot2); library(lattice); library(caret)
library(rpart); library(randomForest)

if(!file.exists('pml-training.csv')){
    download.file("https://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv",
                  destfil="./pml-training.csv")}
if(!file.exists('pml-testing.csv')){
    download.file("https://d396qusza40orc.cloudfront.net/predmachlearn/pml-testing.csv",
                  destfil="./pml-testing.csv")}

train <- read.csv('pml-training.csv', na.strings = c('NA','NAN','#DIV/0!','NaN',''))
test <- read.csv('pml-testing.csv', na.strings=c('NA','NAN','#DIV/0!','NaN',''))

set.seed(12345)

I’m only going to keep columns with mostly non-NA information (more than 50%). As we’ll see, this actually accounts for all NA values in both datasets. Note that I am selecting the columns for both train and test, based on columns only from the train data to keep my matrices consistent.

test <- test[colMeans(!is.na(train))>0.5]
train <- train[colMeans(!is.na(train))>0.5]

(sum(is.na(train))==0) & (sum(is.na(test))==0)

## [1] TRUE

names(train)

##  [1] "X"                    "user_name"            "raw_timestamp_part_1"
##  [4] "raw_timestamp_part_2" "cvtd_timestamp"       "new_window"          
##  [7] "num_window"           "roll_belt"            "pitch_belt"          
## [10] "yaw_belt"             "total_accel_belt"     "gyros_belt_x"        
## [13] "gyros_belt_y"         "gyros_belt_z"         "accel_belt_x"        
## [16] "accel_belt_y"         "accel_belt_z"         "magnet_belt_x"       
## [19] "magnet_belt_y"        "magnet_belt_z"        "roll_arm"            
## [22] "pitch_arm"            "yaw_arm"              "total_accel_arm"     
## [25] "gyros_arm_x"          "gyros_arm_y"          "gyros_arm_z"         
## [28] "accel_arm_x"          "accel_arm_y"          "accel_arm_z"         
## [31] "magnet_arm_x"         "magnet_arm_y"         "magnet_arm_z"        
## [34] "roll_dumbbell"        "pitch_dumbbell"       "yaw_dumbbell"        
## [37] "total_accel_dumbbell" "gyros_dumbbell_x"     "gyros_dumbbell_y"    
## [40] "gyros_dumbbell_z"     "accel_dumbbell_x"     "accel_dumbbell_y"    
## [43] "accel_dumbbell_z"     "magnet_dumbbell_x"    "magnet_dumbbell_y"   
## [46] "magnet_dumbbell_z"    "roll_forearm"         "pitch_forearm"       
## [49] "yaw_forearm"          "total_accel_forearm"  "gyros_forearm_x"     
## [52] "gyros_forearm_y"      "gyros_forearm_z"      "accel_forearm_x"     
## [55] "accel_forearm_y"      "accel_forearm_z"      "magnet_forearm_x"    
## [58] "magnet_forearm_y"     "magnet_forearm_z"     "classe"

Since our primary question is to consider the excercises and their execution, I will ignore timestamp data (including window data) and any data concerning the subject’s identity.

test <- test[,8:ncol(train)]
train <- train[,8:ncol(train)]

Lastly, to help us select which model to use, let’s split the train dataset into training and cross-validation sets (train and cv, respectively).

cvPartition <- createDataPartition(y=train$classe,
                                   p=0.75,
                                   list=FALSE)

cv <- train[-cvPartition,]
train <- train[cvPartition,]

dim(train); dim(cv); dim(test)

## [1] 14718    53

## [1] 4904   53

## [1] 20 53

Training and Cross-Validating

We’ll be training a decision-tree model as well as a random-forests model.

model.DecisionTree <- rpart(classe~., data=train, method='class')
model.RandomForests <- randomForest(classe~., data=train, na.action=na.omit)

Next, we’ll use these models to predict the classifications for the cv set. After that, we’ll compare the models using confusion matrices.

prediction.DecisionTree <- predict(model.DecisionTree,
                                   cv, type='class')

prediction.RandomForests <- predict(model.RandomForests,
                                    cv, type='class')

confusion.DecisionTree <- confusionMatrix(prediction.DecisionTree, cv$classe)
confusion.RandomForests <- confusionMatrix(prediction.RandomForests, cv$classe)

Comparing Models

Below are the comparisons of how well the two models can predict the classe variable of the cv dataframe.

df.DecisionTree <- as.data.frame(confusion.DecisionTree$table)
df.RandomForests <- as.data.frame(confusion.RandomForests$table)

NormalizeFrequency <- function(Iter,Dataframe){
    Dataframe[Iter,'Freq']/sum(subset(Dataframe, Reference==Dataframe[Iter,'Reference'])$Freq)
}

df.DecisionTree$NormFreq <- sapply(1:nrow(df.DecisionTree), 
                                   function(x) NormalizeFrequency(x,df.DecisionTree))
df.RandomForests$NormFreq <- sapply(1:nrow(df.RandomForests), 
                                    function(x) NormalizeFrequency(x,df.RandomForests))

df.DecisionTree$Model <- rep('Decision Tree',nrow(df.DecisionTree))
df.RandomForests$Model <- rep('Random Forests',nrow(df.RandomForests))

df.Models <- rbind(df.DecisionTree, df.RandomForests)

ggplot(data=df.Models, aes(x=Prediction, y=Reference, fill=NormFreq)) + 
    geom_tile() + 
    facet_grid(~Model) +
    guides(fill=FALSE) +
    geom_text(aes(x=Prediction, y=Reference,
                  label=round(NormFreq,3)))

Clearly, the model using Random Forests is better able to predict for all 5 classifications. As such, we will proceed by predicting and comparing the test set.

Predicting the Test Set

Using the same procedure as before and our model.RandomForests model, we can then make predictions for the test set classifications:

finalPrediction <- predict(model.RandomForests, 
                           test, type='class')

finalPrediction

##  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 
##  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

Acknowledgements

I’d like to thank Ugulino, W.; Cardador, D.; Vega, K.; Velloso, E.; Milidiu, R.; and Fuks, H.; all of whom made this dataset and analysis possible.

Read more about their study here: http://groupware.les.inf.puc-rio.br/har#ixzz4j9Jb8hbi.

Prediction Assignment