courseproj.Rmd

---
title: "PML course project"
author: "Dmitry Sherbina"
date: "Saturday, October 25, 2014"
output: html_document
---

load data and libraries
```{r}
library(caret); library(nlme); library(class)
D=read.csv('pml-training.csv')
T=read.csv('pml-testing.csv')
names(D)
dim(D)
```

there are many variables (160) in data, some are strange: picth 

```{r, echo=FALSE}
plot(D$user_name)
```

data of 6 users are relatively evenly distributed

```{r}
qplot(X,classe, data = D, color=D$user_name)
```

samples in dataset are groupped by classe and user_name, but not strictly regular. Random sampling is ok for further slicing.


we want to predict one of five possible outcome in variable "classe"

```{r}
plot(D$classe)
```

Classe A is most probable of them, but comparable, i.e. of the same order


##Preprocessing

change categorical output into numbers

```{r}
Out=predict(dummyVars(~classe, data=D),newdata=D)
head(Out)
```

some variables seem not much useful for prediction, such X or raw.timestamp.part.1
```{r}
N=subset(D, select = -c(X, user_name, cvtd_timestamp, raw_timestamp_part_1, raw_timestamp_part_2, cvtd_timestamp,  new_window,  num_window, classe))
#N=data.frame(N,Out)
```

many variables are dependent of each other, for example space coordinates of body points, e.g. as seen for arm magnet
```{r}
qplot(magnet_arm_x,magnet_arm_z, data=D, color=classe)
sum(complete.cases(N))/ dim(N)[1]
```
also only a fifth of samples are complete

so some dimension reduction will be useful, also remove near zero variated vars

```{r}
N=subset(N, select=!is.na(N[1,]))
nzv = nearZeroVar(N,saveMetrics=TRUE)
N=subset(N, select=!nzv$nzv)
```

to pick variables to include in a model we will use cross-validation later


now transform data to compensate outliers
```{r, echo=FALSE}
pp=preProcess(N, method=c('pca'), thres=0.9, verbose = 0)
NN=predict(pp,N)
#dim(NN)
qplot(PC1,PC2, data=NN)
#qplot(ICA1,ICA2, data=NN, color=classe)
```
so data can be converted to 19 components which explain for 0.9 of variation

## Training

use training method - rpart - Tree-Based Model with cross-validation.
This model is suited for categorical output without converting to indicator vars.

methods failed: rf, svm, WM, treebag, oblique.tree, evtree spls gpls svmSpectrumString  svmRadialWeights svmRadialCost  svmPoly earth bagEarth gcvEarth rpart  bagEarth   ORFpls rpartCost partDSA gamSpline gamboost glmboost blackboost                 , not suited: glm,rlm, plsRglm, lmStepAIC,lm

  for 12 ICA:
  bayesglm(.3) widekernelpls(.45)
  
  methods tried for 12 PC: (Accuracy)
  pls(.37) widekernelpls(.37) kernelpls(.37) rpart(.5) bayesglm(.32) simpls(.37)
  ?ctree (0.57) plr(.32 "Penalized Logistic Regression")      
```{r}
set.seed(555)
classe=D$classe
mo <- train(classe ~.,data=data.frame(N,classe), method="rpart", trControl = trainControl(method='cv')) #, preProcess=c('pca') )
mo
```

```{r}
mo$finalModel
```
not able readily interpret residuals because of PCA, so the final version is w/o preprocessing


## Check

```{r}
R=predict(mo, T)
R
```


```{r}
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(R)
```