Obtaining data

Training and test set was downloaded into working directory. Then empty columns were excluded. Actually it could be done in more elegant way during loading csv files.

train <- read.csv("pml-training.csv", stringsAsFactors = FALSE)
test <- read.csv("pml-testing.csv", stringsAsFactors = FALSE)
minusCol1 <- colSums(train == "") > 10000
minusCol1[is.na(minusCol1)]  <- TRUE
minusCol2 <- colSums(is.na(train)) > 10000
minusCol <- minusCol1 | minusCol2
minusCol[c(1, 3:6)] <- TRUE
train <- train[,!minusCol]

Using user_name variable

I know some students have opinions that it's not necessary to use ID variable. I believe different persons can do things in different ways. That's why I didn't omit this variable. Here's a plot that shows my motivation. For some users parameters may be significant different.

library(ggplot2)
qplot(train[,1], train[,2], colour = train$classe)

Gradient boosting method

Gradient boosting method showed excellent results but it took pretty long time to train model (more than 30 minutes). Accuracy on training set is higher 99%. This model gave 20/20 answers on final test.

library(caret)
# fit_gbm <- train(classe ~., data = tr, method = "gbm")
fit_gbm <- readRDS("gbm01.rds")
confusionMatrix(predict(fit_gbm, train), train$classe)

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction    A    B    C    D    E
##          A 5575   28    0    0    0
##          B    5 3744   17    7    6
##          C    0   25 3399   26    4
##          D    0    0    4 3183   29
##          E    0    0    2    0 3568
## 
## Overall Statistics
##                                           
##                Accuracy : 0.9922          
##                  95% CI : (0.9909, 0.9934)
##     No Information Rate : 0.2844          
##     P-Value [Acc > NIR] : < 2.2e-16       
##                                           
##                   Kappa : 0.9901          
##  Mcnemar's Test P-Value : NA              
## 
## Statistics by Class:
## 
##                      Class: A Class: B Class: C Class: D Class: E
## Sensitivity            0.9991   0.9860   0.9933   0.9897   0.9892
## Specificity            0.9980   0.9978   0.9966   0.9980   0.9999
## Pos Pred Value         0.9950   0.9907   0.9841   0.9897   0.9994
## Neg Pred Value         0.9996   0.9967   0.9986   0.9980   0.9976
## Prevalence             0.2844   0.1935   0.1744   0.1639   0.1838
## Detection Rate         0.2841   0.1908   0.1732   0.1622   0.1818
## Detection Prevalence   0.2855   0.1926   0.1760   0.1639   0.1819
## Balanced Accuracy      0.9986   0.9919   0.9949   0.9939   0.9945

Random forests

Default settings of training model using random forests causing leads to long training time. Using cross validation tuning it's possible to reduce training time. Parameter number divides training set on 2 folds in my case (while default settings have 25 resampling iterations). This gives 100% accuracy on training set and 20/20 on test set. Training time was about 6 minutes and it increases by 5-6 minutes with every new fold.

library(caret)
# set.seed(19)
# fit_rf2 <- train(classe ~., 
#                 method="rf", 
#                 trControl=trainControl(method = "cv",
#                                        number = 2), 
#                 data=train)
fit_rf2 <- readRDS("fit_rf2.rds")
confusionMatrix(predict(fit_rf2, train), train$classe)

## Loading required package: randomForest
## randomForest 4.6-12
## Type rfNews() to see new features/changes/bug fixes.

## Confusion Matrix and Statistics
## 
##           Reference
## Prediction    A    B    C    D    E
##          A 5580    0    0    0    0
##          B    0 3797    0    0    0
##          C    0    0 3422    0    0
##          D    0    0    0 3216    0
##          E    0    0    0    0 3607
## 
## Overall Statistics
##                                      
##                Accuracy : 1          
##                  95% CI : (0.9998, 1)
##     No Information Rate : 0.2844     
##     P-Value [Acc > NIR] : < 2.2e-16  
##                                      
##                   Kappa : 1          
##  Mcnemar's Test P-Value : NA         
## 
## Statistics by Class:
## 
##                      Class: A Class: B Class: C Class: D Class: E
## Sensitivity            1.0000   1.0000   1.0000   1.0000   1.0000
## Specificity            1.0000   1.0000   1.0000   1.0000   1.0000
## Pos Pred Value         1.0000   1.0000   1.0000   1.0000   1.0000
## Neg Pred Value         1.0000   1.0000   1.0000   1.0000   1.0000
## Prevalence             0.2844   0.1935   0.1744   0.1639   0.1838
## Detection Rate         0.2844   0.1935   0.1744   0.1639   0.1838
## Detection Prevalence   0.2844   0.1935   0.1744   0.1639   0.1838
## Balanced Accuracy      1.0000   1.0000   1.0000   1.0000   1.0000

Conclusion

Both methods, Gradient Boosting and Random Forests gave excellent results on training set with accuracy >99% and 100% respectively. Both methods passed final test 20/20. Full code and models available on my repo: https://github.com/yurkai/PML-CP

Ps. To my mind this project was controversial. Because I used really sophisticated and complex methods just with few lines of code without understanding of inner mechanics of these methods. If you'll say that I better spend time on it and there are tons of information you will probably right, that's why took another few courses dedicated machine and statistical learning. Thanks.