ReportProject1.Rmd

---
title: "Project 1 - Practical Machine Learning - Coursera"
author: "Francisco Cano Marchal"
date: "22 de marzo de 2015"
output: html_document
---
```{r}
library(caret)
```


Get the data   

```{r}
url <- "http://d396qusza40orc.cloudfront.net/predmachlearn/pml-training.csv"
download.file(url, destfile="/Users/franciscocanomarchal/Data-Science/Practical-Machine-Learning/Project-One/training.csv", method="curl")
exercise <- read.csv("~/Data-Science/Practical-Machine-Learning/Project-One/training.csv")
```

Divide into training and test set for cross validation   

```{r}
trainPart <- createDataPartition(exercise$classe,p=0.75,list=FALSE)
training <- exercise[trainPart,]
test <- exercise[-trainPart,]
```

Select variables   

```{r}
train <- subset(training,select=-c(X,user_name,raw_timestamp_part_1,raw_timestamp_part_2,cvtd_timestamp, new_window, num_window))    
```

I am going to keep only the variables actually measured   

```{r}
trainm <- train[,grepl(names(train),pattern="^roll_|^pitch_|^yaw_|^gyros_|^accel_|^magnet_|^classe")]
```

```{r}
M <- abs(cor(subset(trainm,select=-c(classe))))
diag(M) <- 0
which(M>0.8,arr.ind=T)
```

Some of them are quite correlated, it is to be expected for ex. due to contraints in the body movement   

I'll limit my descriptors using PCA   

```{r}
traindes <- subset(trainm,select=-c(classe))
        
prComp <- preProcess(traindes,method="pca")
trainPC  <- predict(prComp,traindes)
```

Only 23 components needed to capture 95 percent of the variance   


I'm training my model using RANDOM FORESTS with the previously selected variables   

```{r}
set.seed(647)
library(randomForest)
modelFitPCA <- randomForest(trainm$classe~.,data=trainPC)
```

```{r}
modelFitPCA
```

The model looks good, but still we have to use cross validation to get a more real estimation of the accuracy of model.    

Let's test the model with my testing set   

```{r}
test <- subset(test,select=-c(X,user_name,raw_timestamp_part_1,raw_timestamp_part_2,cvtd_timestamp, new_window, num_window))    
testm <- test[,grepl(names(test),pattern="^roll_|^pitch_|^yaw_|^gyros_|^accel_|^magnet_|^classe")]
testdes <- subset(testm,select=-c(classe))
testPCA <- predict(prComp, testdes)

pred  <- predict(modelFitPCA,testPCA)
table(pred,testm$classe)
confusionMatrix(testm$classe,pred)
```

Not bad.