code_submission.Rmd

---
title: "Practical Machine Learning Course Project"
author: "Alessandro Capuani"
date: "21/04/2015"
output: html_document
---



Train and test data sets were loaded and pre-processed to remove unuseful predictors, e.g. only containing NA or with zero variance.
The variable
```{r}
library(caret)
data_r <- read.table('pml-training.csv',header=T,row.names=1,sep=",")
```
The variables "user_name","raw_timestamp_part_1","raw_timestamp_part_2" and "cvtd_timestamp" were also removed, since they do not have any
influence on the response variable "classe", as shown below.
```{r}
featurePlot(x=data_r[,c("user_name","raw_timestamp_part_1",
                       "raw_timestamp_part_2",
                       "cvtd_timestamp","classe")],y=data_r$classe,
                        plot="pairs")
```


The function nearZeroVar from the caret package was used to detect variables with zero or near zero variance. Finally, after pre-processing
53 predictors were retained for train and test data sets, without considering the response variable.
```{r}
data_filtered <- data_r[,colSums(is.na(data_r)) == 0]
nzv <- nearZeroVar(data_filtered, saveMetrics= FALSE)
data_filtered <- data_filtered[,-nzv]
data_filtered <- data_filtered[,-c(1:4)]

test <- read.table('pml-testing.csv',header=T,row.names=1,sep=",")
test_filtered <-test[,colnames(data_filtered[,1:53])]
dim(data_filtered)
dim(test_filtered)
```

The train data set was splitted in two parts, training and validation data sets. The latest will be used to validate our model.
```{r}
trainIndex <- createDataPartition(data_filtered$classe, p = .8,
                                  list = FALSE,
                                  times = 1)

training <-data_filtered[trainIndex,]
validation <-data_filtered[-trainIndex,]
```

Random forest was use as model due to its high accuracy. A cross validation with 3 folds was applied as resampling method. A seed was set because of the randomization reproducibility. The trained model showed a very high accuracy, as also displayed in the output below.
```{r}
set.seed(3254)
fitControl <- trainControl(## 3-fold CV
                           method = "cv",
                           number = 3,
                           ## repeated ten times
                           repeats = 3,
                           verboseIter = FALSE)

modelFit_RF <-train(classe ~.,data=training,method="rf",trControl = fitControl,verbose=T)
modelFit_RF
```
Moreover, also using the validation set, it is possible to see that the error within the categories very low is.
```{r}
pred <-predict(modelFit_RF,validation[,1:53])
known <-validation$classe
table(pred,known)
```
Thanks to wugulino at inf dot puc-rio dot br to share their data set-