In andreabecsek/NaiveBayes: Gaussian Naive Bayes model

knitr::opts_chunk$set(echo = TRUE)

library(tidyverse)
library(ggplot2)
library(magrittr)
library(dplyr)

data <- read.csv("~/Documents/Project/stats-computing-1/data/data.csv")
data %<>%  
  dplyr::select(-c(id, X))

# use 70% of the data for training and 30% for testing
n <- nrow(data)
train_size <- floor(0.8*n)

# sample indeces for split
train_ind <- sample(seq(n),size = train_size, replace = FALSE)
train <- data[train_ind,]
test <- data[-train_ind,]

y <- train$diagnosis
X <- train %>% select(-diagnosis) %>% as.matrix()

y_test <- test$diagnosis
X_test <- test %>% select(-diagnosis) %>% as.matrix()

X <- unname(cbind(1,X))
X_test <- unname(cbind(1,X_test))
# change factor levels to B=0 and M=1
levels(y) <- c(0,1)
y %<>% matrix
mode(y) <- 'numeric'

levels(y_test) <-c(0,1)
y_test %<>% matrix
mode(y_test) <- 'numeric'

Calculate mean and variance for for every feature for both classes

```r

# create empty matrices for all classes

summaries <- function(X,y){

d <- dim(X)[2]

k <- length(unique(y))

summ <- matrix(rep(0,d*2),ncol=d)

summs <- replicate(2,summ,simplify = F)

for(i in 1:k){

summs[[i]] <- t(sapply(1:d, function(x) c(mean(X[,x]),sd(X[,x])),simplify = T))

}

return(summs)

}

```

Class probabilities

prior <- function(y){
  k <- sort(unique(y)[,1])
  prior <- rep(0,length(k))
  for(i in 1:length(k)){
    prior[i] <- sum(y==k[i])/length(y)
  }
  return(prior)
}

prior(y)

# create empty matrices for all classes
summaries <- function(X,y){
  d <- dim(X)[2]
  k <- length(unique(y))
  classes <- array(rep(1,d*k*2), dim=c(k,d,2))
  for(i in 1:k){
    X_k <- X[which(y==(i-1)),]
    classes[i,,1] <- apply(X_k,2,mean)
    classes[i,,2] <- apply(X_k,2,sd)
  }
  return(classes)
}

Conditional probs p(x/class)

# for x_new a single observation
cond <- function(X_new,summaries,priors){
  n <- dim(X_new)[1]
  d <- dim(X_new)[2]
  k <- dim(summaries)[1]
  result <- matrix(rep(0,n*k),nrow=n)
  for(obs in 1:n){
    for(class in 1:k){
      post <- log(priors[class])
      for(feat in 1:d){
        mu <- summaries[class,feat,1]
        sd <- summaries[class,feat,2]
        cond <- dnorm(X_new[obs,feat],mu, sd, log = TRUE)
        post <- post + cond
      }
      result[obs,class] <- post
    }
  }
  pred <- apply(result,1,which.max) 
  # return(result)
  return(as.matrix(pred-1))
}  

confusion_matrix <- as.data.frame(table(predictions,y_test))

ggplot(confusion_matrix,aes(y=predictions,x=y_test))+
  geom_tile(aes(fill=Freq))+
  geom_text(aes(label=sprintf("%1.0f", Freq)),color="white",fontface="bold")+
  labs(y="Predicted class",x="True class")+
  theme_minimal()

andreabecsek/NaiveBayes documentation built on Jan. 19, 2020, 12:49 p.m.