index.Rmd

---
title: "Practical Machine Learning Course Project"
author: "Yuya Katafuchi"
date: "16 November 2018"
output: html_document
---

## Objective
For this course project, we aim to predict the manner in exercise given the dataset consisting of a large amount of personal data obtained by accelerometers on the belt, forearm, arm and dumbell of six participants.

## Load Package
We load the caret package for the data preparation and analysis, the tidyverse package for data cleansing, and the readr package for the data loading.
```{r}
library(caret)
library(readr)
library(tidyverse)
```


## Load Data

First of all, we load training and testing dataset.

```{r message = FALSE, warning = FALSE}
training <- read_csv("pml-training.csv")
testing <- read_csv("pml-testing.csv")
as.tibble(training)
```

Our target variable to predict is "classe", which is the performance of exercise and classified as 4 levels.
```{r}
training <- training %>% mutate(classe = factor(classe))
table(training$classe)
```

## Cleansing Data
After looking the dataset structure, we found that some of columns is not useful for our prediction.
```{r}
training %>% select(X1, user_name, raw_timestamp_part_1, 
                    raw_timestamp_part_2, cvtd_timestamp,
                    new_window, num_window)
```
We then remove these columns for the accurate prediction.
```{r}
training_2 <- training %>% select(-X1, -user_name, -raw_timestamp_part_1,
                                  -raw_timestamp_part_2, -cvtd_timestamp,
                                  -new_window, -num_window)
```
As well as useless columns, we need to determinate which column as almost same features by cheking near-zero-variance.
```{r}
chk_variance <- nearZeroVar(training_2, saveMetrics = TRUE)
head(chk_variance)
```
We then remove these TRUE columns by:
```{r}
training_3 <- training_2[, !chk_variance$nzv]
```
Again, we remove the useless column consisting NAs and blanks.
```{r}
NAcol <- training_3 %>% sapply(is.na) %>% apply(2, sum)
training_df <- training_3[, NAcol == 0]
training_df
```
After data cleansing, we have 49 features to predict classe.

## Split Data
To measure the out-of-sample error without possibilities to be overfitting, we create cross-validation set (20%) by:
```{r}
set.seed(145)
training_split <- createDataPartition(training_df$classe, p = 0.8, list = FALSE)
training_f <- training_df[training_split, ]
cv_f <- training_df[-training_split, ]
```

## Data Cleansing for Testing Set
We do the same cleansing as training set for test set.
```{r}
testing_2 <- testing %>% select(-X1, -user_name, -raw_timestamp_part_1,
                                  -raw_timestamp_part_2, -cvtd_timestamp,
                                  -new_window, -num_window)
testing_3 <- testing_2[, !chk_variance$nzv]
testing_f <- testing_3[, NAcol == 0]

```


## Learn Model
Again, prediction target variable is "classe", which is the performance of exercise and classified as 4 levels. We then decide to use random forest, because of its easiness and power of multiple classification. To begin with, the model includes all of predictors in final dataset.
```{r warning=FALSE}
rf_full <- train(classe ~., data = training_f, method = "rf")
rf_full$finalModel
```

## Less Features
Second, we will estimate simpler model omitting location information with _x, _y, _z.
```{r warning=FALSE}
training_s_f <- training_f %>% select(roll_belt, roll_arm, roll_dumbbell, roll_forearm, 
                                      pitch_belt, pitch_arm, pitch_dumbbell, pitch_forearm,
                                      yaw_belt, yaw_arm, yaw_dumbbell, yaw_forearm,
                                      total_accel_belt, total_accel_arm, total_accel_dumbbell,
                                      total_accel_forearm, classe)
rf_simple <- train(classe ~., data = training_s_f, method = "rf")
rf_simple$finalModel
```

##Cross-Validation test
Using cross validation set we made, we will check the prediction accuracy without taking look at test set.
```{r}
predict_full <- predict(rf_full, cv_f)
predict_simple <- predict(rf_simple, cv_f)
confusionMatrix(predict_full, cv_f$classe)
confusionMatrix(predict_simple, cv_f$classe)
```

##Predict using test set
Judging from above confusion matrix, we chose simpler specification to predict the test classe since it has reasonably high accuracy even though it contains less predictors.
```{r}
predict(rf_simple, testing_f)
```