Lachlan Miller
May 4, 2018 · 7 min read
Image for post
Image for post
Sample data and several hypothesis’ for this article

Understanding and Calculating the Cost Function for Linear Regression

This post will focus on the properties and application of cost functions, how to solve it them by hand. Then we will implement the calculations twice in Python, once with for loops, and once with vectors using numpy. This goes into more detail than my previous article about linear regression, which was more a high level summary of the concepts.

When learning about linear regression in Andrew Ng’s Coursera course, two functions are introduced:

  • the cost function

At first I had trouble understanding what each was for. Together they form linear regression, probably the most used learning algorithm in machine learning.

What is a Cost Function?

In the case of gradient descent, the objective is to find a line of best fit for some given inputs, or X values, and any number of Y values, or outputs. A cost function is defined as:

…a function that maps an event or values of one or more variables onto a real number intuitively representing some “cost” associated with the event.

from Wikipedia

In this situation, the event we are finding the cost of is the difference between estimated values, or the hypothesis and the real values — the actual data we are trying to fit a line to.

To state this more concretely, here is some data and a graph.

║ X ║ y ║
║ 1.00 ║ 1.00 ║
║ 2.00 ║ 2.50 ║
║ 3.00 ║ 3.50 ║
Image for post
Image for post

Pretty boring graph. The goal here is to find a line of best fit — a line that approximates the values most accurately. Here are some random guesses:

Image for post
Image for post
Some random guesses at a line of best fit
║ X ║ y ║ best_fit_1 ║ best_fit_2 ║ best_fit_3 ║
║ 1.00 ║ 1.00 ║ 0.50 ║ 1.00 ║ 1.50 ║
║ 2.00 ║ 2.50 ║ 1.00 ║ 2.00 ║ 3.00 ║
║ 3.00 ║ 3.50 ║ 1.50 ║ 3.00 ║ 4.00 ║

Making that beautiful table was really hard, I wish Medium supported tables. Anyway, we have three hypothesis — three potential sets of data that might represent a line of best fit. The slope for each line is as follows:

  • best_fit_1: 0.5

best_fit_2 looks pretty good , I guess. But we are data scientists, we don’t guess, we conduct analysis and make well founded statements using mathematics.

Understanding the Cost Function

Let’s do an analysis using the squared error cost function.

Image for post
Image for post
Squared Error cost function

Remember a cost function maps event or values of one or more variables onto a real number. In this case, the event we are finding the cost of is the difference between estimated values, or the difference between the hypothesis and the real values — the actual data we are trying to fit a line to.

Let’s unwrap the mess of greek symbols above. On the far left, we have 1/2*m. m is the number of samples — in this case, we have three samples for X. Those are 1, 2 and 3. So 1/2*m is a constant. It turns out to be 1/6, or 0.1667 .

Next we have a sigma.

Image for post
Image for post

This means the sum. In this case, the sum from i to m, or 1 to 3. We repeat the calculation to the right of the sigma, that is:

Image for post
Image for post

for each sample.

The actual calculation is just the hypothesis value for h(x), minus the actual value of y. Then you square whatever you get.

The final result will be a single number. We repeat this process for all the hypothesis, in this case best_fit_1 , best_fit_2 and best_fit_3. Whichever has the lowest result, or the lowest “cost” is the best fit of the three hypothesis.

Let’s go ahead and see this in action to get a better intuition for what’s happening.

Calculating the Cost Function by Hand

Let’s run through the calculation for best_fit_1.

║ X ║ y ║ best_fit_1 ║
║ 1.00 ║ 1.00 ║ 0.50 ║
║ 2.00 ║ 2.50 ║ 1.00 ║
║ 3.00 ║ 3.50 ║ 1.50 ║

We know the 1/2m is simply 1/6, so we will focus on the summing the result of:

Image for post
Image for post

for best_fit_1, where i = 1, or the first sample, the hypothesis is 0.50. This is the h_theha(x(i)) part, or what we think is the correct value. The actual value for the sample data is 1.00. So we are left with (0.50 — 1.00)^2 , which is 0.25. Let’s add this result to an array called results.

results = [0.25]

The next sample is X = 2. The hypothesis value is 1.00 , and the actual y value is 2.50 . So we get (1.00 — 2.50)^2, which is 2.25. Add it to results.

results = [0.25, 2.25]

Lastly, for X = 3, we get (1.50 — 3.50)^2 , which is 4.00.

results = [0.25, 2.25, 4.00]

Finally, we add them all up and multiply by 1/6.

6.5 * (1/6) = 1.083. The cost is 1.083. That’s nice to know, but we need some more costs to compare it to. Go ahead and repeat the same process for best_fit_2 and best_fit_3. I already did it, and I got:

best_fit_1: 1.083
best_fit_2: 0.083
best_fit_3: 0.25

A lowest cost is desirable. A low costs represents a smaller difference. By minimizing the cost, we are finding the best fit. Out of the three hypothesis presented, best_fit_2 has the lowest cost. Reviewing the graph again:

Image for post
Image for post
best_fit_2 has the lowest cost

The orange line, best_fit_2, is the best fit of the three. We can see this is likely the case by visual inspection, but now we have a more defined process for confirming our observations.

Use with Linear Regression

We just tried three random hypothesis — it is entirely possible another one that we did not try has a lowest cost than best_fit_2. This is where Gradient Descent (henceforce GD) comes in useful. We can use GD to find the minimized value automatically, without trying a bunch of hypothesis one by one. Using the cost function in in conjunction with GD is called linear regression.

This will be the topic of a future post. For now, I want to focus on implementing the above calculations using Python.

Calculating the cost function using Python (#1)

As promised, we perform the above calculations twice with Python. Once using for loops, and once using vectors. Firstly, with for loops. The focus of this article is the cost function, not how to program Python, so the code is intentionally verbose and has lots of comments to explain what’s going on.

Personally, the biggest challenge I am facing is how to take the theoretically knowledge and algorithms I learned in my undergraduate calculus classes (I studied electrical engineering) and turn them into working code.

The way I am breaking this barrier down is by really understanding what is going on when I see a equation on paper, and once I understand it (usually after doing several iterations by hand), it’s lot easier to turn into code. Hopefully this helps other people, too.

cost function calculations using for loops

It is more common to perform the calculations “all at once” by turning the data set and hypothesis into matrices. This process is called vectorization.

Calculating the cost function using Python (#2)

It’s a little unintuitive at first, but once you get used to performing calculations with vectors and matrices instead of for loops, your code will be much more concise and efficient. Here is the same calculation implemented with matrices using numpy.

calculating cost function with vectors

Even without reading the code, it’s a lot more concise and clean.

We are using numpy, and defining X and y as np.array. They provide many more properties for doing vector and matrices multiplication. There are two things to note:

  1. We prepend the X vector with a vector 1 s. This is explained well by Andrew Ng here, in week 1’s “matrix matrix multiplication” lecture. Basically, when you multiply matrices, you need to have the correct dimensions. The vector of 1s allows for this. This is best explained using video, I recommend you watch it on Coursera.
inner = np.power(((X @ theta.T) - y, 2)

Which is doing this:

Image for post
Image for post
the ‘inner’ calculation


np.sum(inner) / 2 * len(X)

which is doing

Image for post
Image for post

Again, I encourage you to sign up for the course (it’s free) and watch the lectures under week 1’s “linear algebra review”. It takes a while to really get a feel for this style of calculation. I went through and put a ton of print statements, and inspected the contents of the array, as well as the array.shape property to really understand what was happening.


Researching and writing this really solidified by understanding of cost functions. I hope to write a follow up post explaining how to implement gradient descent, first by hand then using Python.

Welcome to a place where words matter. On Medium, smart voices and original ideas take center stage - with no ads in sight. Watch
Follow all the topics you care about, and we’ll deliver the best stories for you to your homepage and inbox. Explore
Get unlimited access to the best stories on Medium — and support writers while you’re at it. Just $5/month. Upgrade

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store