R中的多元线性回归-梯度下降

Question

I am learning machine learning. So I do some simple practice with the data I find online. Right now I try to implement linear regression by gradient descent in R. When I run it, I realize that it does not converge and my cost goes high infinitely. Although I suspect it is somewhere in the part where I calculate gradient, I am not able to find the problem. So lets start presenting my data.

• Dataset : dataset_multipleRegression.csv

My data set contains 4 column : ROLL ~ UNEM, HGRAD, INC So, the goal is finding relationship between ROLL and others.

• Let me present my code

   datavar <- read.csv("dataset.csv") attach(datavar) X <- cbind(rep(1, 29), UNEM,HGRAD,INC) y <- ROLL # function where I calculate my prediction h <- function(X, theta){ return(t(theta) %*% X) } # function where I calculate the cost with current values cost <- function(X, y, theta){ result <- sum((X %*% theta - y)^2 ) / (2*length(y)) return(result) } # here I calculate the gradient, #mathematically speaking I calculate derivetive of cost function at given points gradient <- function(X, y, theta){ m <- nrow(X) sum <- c(0,0,0,0) for (i in 1 : m) { sum <- sum + (h(X[i,], theta) - y[i]) * X[i,] } return(sum) } # The main algorithm gradientDescent <- function(X, y, maxit){ alpha <- 0.005 m <- nrow(X) theta <- c(0,0,0,0) cost_history <- rep(0,maxit) for (i in 1 : maxit) { theta <- theta - alpha*(1/m)*gradient(X, y, theta) cost_history[i] <- cost(X, y, theta) } plot(1:maxit, cost_history, type = 'l') return(theta) } 

I run the code like this

 gradientDescent(X, y, 20)

This is the output I get :

-7.001406e+118  -5.427330e+119  -1.192040e+123  -1.956518e+122

So, can you find where I was wrong. I have already tried different alpha values, didn't make a difference. By the way, I appreciate any tips or good practice from you,

Thanks

Answer 1

Well, I think I finally found the answer. The problem was that I did not appy any feature scaling. Couse I though it was optional precedure for running the algorithm smoothly. Now it works as expected. You can try to run code with scaled dataset using R's scale() function.