pinakbhusan / practical-machine-learning

Predicting the Quantified Self

Bastiaan Quast
Monday, July 21, 2014

Abstract

We use the random forest method to estimate features for the Human Activity Recognition data set from Groupware. We that this method produces relevant results.

Data

The data is taken from the Human Activity Recognition programme at Groupware.

We start the data loading procedure by specifying the data sources and destinations.

training.file <- 'pml-training.csv'
test.file     <- 'pml-test.csv'
training.url  <- 'https://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv'
test.url      <- 'https://d396qusza40orc.cloudfront.net/predmachlearn/pml-testing.csv'

We then execute the downloads.

download.file(training.url, training.file, method='wget')
download.file(test.url,     test.file,     method='wget')

In order to ensure all transformations are applied to both the trainig and test data sets equally, we employ the OOP method and create functions for every step. First we create a function to read data and set NAs and apply this to both datasets.

read.pml       <- function(x) { read.csv(x, na.strings = c("", "NA", "#DIV/0!") ) }
training       <- read.pml(training.file)
test           <- read.pml(test.file)
training       <- training[,-c(1,5,6)]
test           <- test[,-c(1,5,6)]

library(caret)
trainingIndex  <- createDataPartition(training$classe, p=.50, list=FALSE)
training.train <- training[ trainingIndex,]
training.test  <- training[-trainingIndex,]

Next we create a function to remove entire NA columns, and apply it to both data frames. Lastly we create a function that removes any variables with missing NAs and apply this to the training set.

rm.na.cols     <- function(x) { x[ , colSums( is.na(x) ) < nrow(x) ] }
training.train <- rm.na.cols(training.train)
training.test  <- rm.na.cols(training.test)
complete       <- function(x) {x[,sapply(x, function(y) !any(is.na(y)))] }
incompl        <- function(x) {names( x[,sapply(x, function(y) any(is.na(y)))] ) }
trtr.na.var    <- incompl(training.train)
trts.na.var    <- incompl(training.test)
training.train <- complete(training.train)
training.test  <- complete(training.test)

Method

We use the Random Forests method [@breiman2001random], which applies bagging to tree learners. The Bootstap Aggregating (bagging) method is described in Technical Report No. 421: Bagging Predictors [@breiman1996bagging].

library(randomForest)
random.forest <- train(training.train[,-57],
                       training.train$classe,
                       tuneGrid=data.frame(mtry=3),
                       trControl=trainControl(method="none")
                       )

Results

Some statistics on the results

summary(random.forest)

##                 Length Class      Mode     
## call                4  -none-     call     
## type                1  -none-     character
## predicted        9812  factor     numeric  
## err.rate         3000  -none-     numeric  
## confusion          30  -none-     numeric  
## votes           49060  matrix     numeric  
## oob.times        9812  -none-     numeric  
## classes             5  -none-     character
## importance         56  -none-     numeric  
## importanceSD        0  -none-     NULL     
## localImportance     0  -none-     NULL     
## proximity           0  -none-     NULL     
## ntree               1  -none-     numeric  
## mtry                1  -none-     numeric  
## forest             14  -none-     list     
## y                9812  factor     numeric  
## test                0  -none-     NULL     
## inbag               0  -none-     NULL     
## xNames             56  -none-     character
## problemType         1  -none-     character
## tuneValue           1  data.frame list     
## obsLevels           5  -none-     character

We now compare the results from the predition with the actual data.

confusionMatrix(predict(random.forest,
                        newdata=training.test[,-57]),
                training.test$classe
                )

## Loading required package: randomForest
## randomForest 4.6-7
## Type rfNews() to see new features/changes/bug fixes.

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction    A    B    C    D    E
##          A 2790    5    0    0    0
##          B    0 1893    7    0    0
##          C    0    0 1703    5    0
##          D    0    0    1 1602    2
##          E    0    0    0    1 1801
## 
## Overall Statistics
##                                           
##                Accuracy : 0.9979          
##                  95% CI : (0.9967, 0.9987)
##     No Information Rate : 0.2844          
##     P-Value [Acc > NIR] : < 2.2e-16       
##                                           
##                   Kappa : 0.9973          
##  Mcnemar's Test P-Value : NA              
## 
## Statistics by Class:
## 
##                      Class: A Class: B Class: C Class: D Class: E
## Sensitivity            1.0000   0.9974   0.9953   0.9963   0.9989
## Specificity            0.9993   0.9991   0.9994   0.9996   0.9999
## Pos Pred Value         0.9982   0.9963   0.9971   0.9981   0.9994
## Neg Pred Value         1.0000   0.9994   0.9990   0.9993   0.9998
## Prevalence             0.2844   0.1935   0.1744   0.1639   0.1838
## Detection Rate         0.2844   0.1930   0.1736   0.1633   0.1836
## Detection Prevalence   0.2849   0.1937   0.1741   0.1636   0.1837
## Balanced Accuracy      0.9996   0.9982   0.9974   0.9980   0.9994

The Kappa statistic of 0.994 reflects the out-of-sample error.

plot( varImp(random.forest) )

References