Using devices such as Jawbone Up, Nike FuelBand, and Fitbit it is now possible to collect a large amount of data about personal activity relatively inexpensively. These type of devices are part of the quantified self movement – a group of enthusiasts who take measurements about themselves regularly to improve their health, to find patterns in their behavior, or because they are tech geeks.
One thing that people regularly do is quantify how much of a particular activity they do, but they rarely quantify how well they do it. In this project, our goal is to use data from accelerometers on the belt, forearm, arm, and dumbell of 6 participants. They were asked to perform barbell lifts correctly and incorrectly in 5 different ways. More information is available from the website here (see the section on the Weight Lifting Exercise Dataset). Data for this project has been taken from the same source.
library(caret)
library(rpart)
library(rpart.plot)
library(rattle)
library(randomForest)
trainUrl <- "https://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv"
testUrl <- "https://d396qusza40orc.cloudfront.net/predmachlearn/pml-testing.csv"
readTrain <- read.csv(url(trainUrl))
readTest <- read.csv(url(testUrl))
dim(readTrain)
## [1] 19622 160
dim(readTest)
## [1] 20 160
We see that the training data set has 19622 records and the testing data set has 20 records. The number of variables is 160.
var0 <- nearZeroVar(readTrain)
train <- readTrain[,-var0]
test <- readTest[,-var0]
dim(train)
## [1] 19622 100
We see that 60 redundant variables are removed in the first step.
valNA <- sapply(train, function(x) mean(is.na(x))) > 0.95
train <- train[, valNA == FALSE]
test <- test[, valNA == FALSE]
dim(train)
## [1] 19622 59
The second step leaves 59 variables.
train <- train[,8:59]
test <- test[,8:59]
We now take a look at the column names of the data set.
colnames(train)
## [1] "pitch_belt" "yaw_belt" "total_accel_belt"
## [4] "gyros_belt_x" "gyros_belt_y" "gyros_belt_z"
## [7] "accel_belt_x" "accel_belt_y" "accel_belt_z"
## [10] "magnet_belt_x" "magnet_belt_y" "magnet_belt_z"
## [13] "roll_arm" "pitch_arm" "yaw_arm"
## [16] "total_accel_arm" "gyros_arm_x" "gyros_arm_y"
## [19] "gyros_arm_z" "accel_arm_x" "accel_arm_y"
## [22] "accel_arm_z" "magnet_arm_x" "magnet_arm_y"
## [25] "magnet_arm_z" "roll_dumbbell" "pitch_dumbbell"
## [28] "yaw_dumbbell" "total_accel_dumbbell" "gyros_dumbbell_x"
## [31] "gyros_dumbbell_y" "gyros_dumbbell_z" "accel_dumbbell_x"
## [34] "accel_dumbbell_y" "accel_dumbbell_z" "magnet_dumbbell_x"
## [37] "magnet_dumbbell_y" "magnet_dumbbell_z" "roll_forearm"
## [40] "pitch_forearm" "yaw_forearm" "total_accel_forearm"
## [43] "gyros_forearm_x" "gyros_forearm_y" "gyros_forearm_z"
## [46] "accel_forearm_x" "accel_forearm_y" "accel_forearm_z"
## [49] "magnet_forearm_x" "magnet_forearm_y" "magnet_forearm_z"
## [52] "classe"
We divide our training data (train) into 2 sets, training (60%) and testing (40%). We will use the original testing data, test as our validation set.
trainClasse <- createDataPartition(train$classe, p=0.6, list=FALSE)
training <- train[trainClasse,]
testing <- train[-trainClasse,]
treeModfit <- train(classe ~ ., data = training, method="rpart")
treePred <- predict(treeModfit, testing)
confusionMatrix(treePred, as.factor(testing$classe))
## Confusion Matrix and Statistics
##
## Reference
## Prediction A B C D E
## A 1674 354 72 122 77
## B 58 684 56 81 302
## C 458 343 1074 689 436
## D 40 136 154 394 7
## E 2 1 12 0 620
##
## Overall Statistics
##
## Accuracy : 0.5667
## 95% CI : (0.5556, 0.5777)
## No Information Rate : 0.2845
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.452
##
## Mcnemar's Test P-Value : < 2.2e-16
##
## Statistics by Class:
##
## Class: A Class: B Class: C Class: D Class: E
## Sensitivity 0.7500 0.45059 0.7851 0.30638 0.42996
## Specificity 0.8887 0.92146 0.7027 0.94863 0.99766
## Pos Pred Value 0.7281 0.57917 0.3580 0.53899 0.97638
## Neg Pred Value 0.8994 0.87487 0.9393 0.87463 0.88601
## Prevalence 0.2845 0.19347 0.1744 0.16391 0.18379
## Detection Rate 0.2134 0.08718 0.1369 0.05022 0.07902
## Detection Prevalence 0.2930 0.15052 0.3824 0.09317 0.08093
## Balanced Accuracy 0.8193 0.68603 0.7439 0.62750 0.71381
rpart.plot(treeModfit$finalModel, roundint=FALSE)
We see that the accuracy ≈ 50%, which is quite low.
forestModfit <- train(classe ~ ., data = training, method = "rf", ntree = 100)
forestPred <- predict(forestModfit, testing)
forestPredConfusion <- confusionMatrix(forestPred, as.factor(testing$classe))
forestPredConfusion
## Confusion Matrix and Statistics
##
## Reference
## Prediction A B C D E
## A 2231 12 0 0 0
## B 0 1502 20 0 0
## C 0 3 1344 14 2
## D 1 0 4 1272 2
## E 0 1 0 0 1438
##
## Overall Statistics
##
## Accuracy : 0.9925
## 95% CI : (0.9903, 0.9943)
## No Information Rate : 0.2845
## P-Value [Acc > NIR] : < 2.2e-16
##
## Kappa : 0.9905
##
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: A Class: B Class: C Class: D Class: E
## Sensitivity 0.9996 0.9895 0.9825 0.9891 0.9972
## Specificity 0.9979 0.9968 0.9971 0.9989 0.9998
## Pos Pred Value 0.9947 0.9869 0.9861 0.9945 0.9993
## Neg Pred Value 0.9998 0.9975 0.9963 0.9979 0.9994
## Prevalence 0.2845 0.1935 0.1744 0.1639 0.1838
## Detection Rate 0.2843 0.1914 0.1713 0.1621 0.1833
## Detection Prevalence 0.2859 0.1940 0.1737 0.1630 0.1834
## Balanced Accuracy 0.9987 0.9931 0.9898 0.9940 0.9985
We see that the accuracy ≈ 99%, which is great. hence, we select the random forest model as our prediction model for this analysis.
We now apply our model to the testing data, test
finalPred <- predict(forestModfit, test )
finalPred
## [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
As we see from the result, the random forest outperforms the decision tree in terms of accuracy. While the decision tree gives us ≈50% accuracy, using the random forest gives us a whooping 99% accuracy.