finalproj.Rmd

Practical Machine Learning Project
========================================================

**20-Jun-2014**

*Summary*

Given my unfamiliarity with R, (as well as R markdown) and a limited amount of time, I elected to perform a simple analysis of the datasets provided:
-  Clean up empty columns
-	Split the training data into regression and cross-validation datasets
-	Select relevant factors
-	Preprocess the data using PCA
-	Fit a RandomForest model to the regression dataset, using internal cross-validation
-	Determine prediction quality on cross-validation portion of the training dataset
-	Apply the model to the testing data

**Clean up dataset**

I used the apply function to remove columns with NA's, as well as manually extracted columns that did not appear to be instrument data.

```{r}
library(AppliedPredictiveModeling)
library(caret)
library(minerva)

tr<-read.csv('pml-training.csv',na.strings=c("NA",""),header=T,stringsAsFactors=F)
te<-read.csv('pml-testing.csv',na.strings=c("NA",""),header=T,stringsAsFactors=F)

# pop out nans
te<-te[!apply(te, 2, function(y) any(is.na(y))) ] 
tr<-tr[!apply(tr, 2, function(y) any(is.na(y))) ] 



```

**Split the training data**

I used the createDataPartition() function to split the data into training (for regression) and test sets (for manual cross-validation).

```{r}

inTrain = createDataPartition(tr$classe, p = 0.7)[[1]]
training = (tr[ inTrain,8:59])# select all but classe
testing = (tr[-inTrain,8:59]) # select all but problem id
realtest<-te[,8:59]

```

**Select relevant factors**

Selecting the relevant factors could have been done my simple linear covariance, but I was interested in applying package I've heard about for scoring non-linear correlations, the Minerva library.  I used the mine() function to determine the correlated parameters.

```{r}
# find correlations
#M <- abs(cor(training))
M <- abs(cor(training)) #M<-mine(training)  #try to use MINE (takes too long!)
diag(M) <- 0
ix<-which(M > 0.8,arr.ind=T)
print(ix)
```

**Preprocess**

I used the preprocess() function, as in the quizzes, to perform PCA on the training dataset.

```{r}
trpp=preProcess(training[,ix],method='pca') # use PCA to preprocess

classe<-tr$classe[inTrain]
trfull<-cbind(predict(trpp,training[,ix]),classe) #set for training and caret crossval
classe<-tr$classe[-inTrain]
tefull<-cbind(predict(trpp,testing[,ix]),classe) # sim testing set
testfull<-predict(trpp,realtest[,ix]) # real testing set


```



**Fit the Model**

I used the RandomForest method (method='rf') to fit the dataset, as in quiz 3.  I also used cross-validation as part of the trControl option.  This provided another level of protection against overfitting.

```{r}

modFit1<-train(classe~.,data=trfull,method='rf',trControl=trainControl(method='cv'),number=3,na.action=na.omit)
#modFit2<-train(classe~.,data=training,method='gbm',verbose=F)


save.image("proj_data.RData")
pr<-predict(modFit1,newdata=trfull)
prte<-predict(modFit1,newdata=tefull)
answers<-predict(modFit1,newdata=testfull)

```

**Determine prediction quality**

To determine the quality of the model, my cross-validation set was used with the confusionMatrix() function to examine the confusion matrix to measure the expected out-of-sample-error.  The accuracy of the model on the cross-validation set was:

```{r}

cm<-confusionMatrix(prte,tefull$classe)# look at quality of fit
print(cm)


```

Finally, save the output files:

```
print("write output")
pml_write_files = function(x){
  n = length(x)
  for(i in 1:n){
    filename = paste0("problem_id_",i,".txt")
    write.table(x[i],file=filename,quote=FALSE,row.names=FALSE,col.names=FALSE)
  }
}

pml_write_files(answers)

```
**Apply to testing data**

The trained random forest model was applied to the blinded testing data, and uploaded to the submission page.  The accuracy was close to that of the cross-validation set.  17/20 (85%) were correct