## Machine Learning using SQL – Day 6/100

The below article is the intellectual property of Ashish Kohli. This is one such article which actually powers the ability of SQL. Give it a read guys.

Yes, you read that one right! One of the most fundamental machine learning algorithms out there is Linear Regression. In simple words, it is a technique to describe a relationship between a response(a.k.a dependent) variable and one or more explanatory (a.k.a independent) variable(s). After doing some reading into the “math” behind these algorithms, I realized that this can be easily achieved in SQL.

I don’t intend to turn this post into another “Machine Learning 101”. There are plenty of such articles out there that explain what Linear Regression is in a much better way, including some nitty-gritty details like it’s back-end optimization algorithm, i.e. Gradient Descent. I will thus try to keep this article as less “ML”ly as possible. However, I’d recommend this 9 byte-sized (typo intended!) article series that explains all this and some more in a really easy language. Also, for most of this post, I will be referring to the formulas and notations used in the hyperlinked article.

Alright! At this point, I hope that you already know about the concepts of Linear Regression and how Gradient Descent works. And thus you’ll also know that the relationship between response and explanatory variable(s) is explained by the following equation :

The goal of Linear Regression is to find the optimal value of θ (theta) that best describes the relationship between two variables and Gradient Descent is the way to do that. The equation given below summarizes how Gradient Descent arrives at the optimal value of θ(s).

Let’s list down the activities that are needed to be performed to arrive at optimal values of θ:

- Start with random values of θ and calculate the value of hypothesis function (Hθ)
- Fill in the values of θ and Hθ in the convergence equation to get new values of θ
- Keep repeating Step 2 until the values of θ don’t change anymore
- These values of θ correspond to the minimum cost (or error) for the Hypothesis function
- Fill in the final values of θ in the hypothesis function to get Predicted values of the response variable

**Step 1: Declaring & Initializing variables**

We need 6 variables, each for a different purpose:

- theta0, theta1 to store the current value of θ0, θ1
- theta0_t, theta1_t to store temporary values of θ0 & θ1 before updating the original ones
- var (short for variability) to check if the updated value of θ is approaching “near” the current value or not
- alpha to store learning rate (read this answer at Quora to understand all about learning rate)

DECLARE @theta0 int;

DECLARE @theta1 int;

DECLARE @theta0_t int;

DECLARE @theta1_t int;

DECLARE RomilVaradkar DECIMAL(4,3);

DECLARE @alpha DECIMAL(4,2);

--Initial values

SET @theta0 = 0;

SET @theta1 = 1;

SET @theta0_t = 1;

SET @theta1_t = 0;

SET @alpha = 0.1;

SET RomilVaradkar = 0.01;

**Step 2: Calculating values of Hθ and updated values of θ**

--Calculating theta0

SELECT

@theta0_t = @theta0 - (SUM(Outp)/(SELECT COUNT(*) FROM base))*@alpha

FROM

(

SELECT

(@theta1*X + @theta0) - Y as Outp

FROM base

);--Calculating theta1

SELECT

@theta1_t = @theta1 - (SUM(Outp)/(SELECT COUNT(*) FROM base))*@alpha

FROM

(

SELECT

((@theta1*X + @theta0) - Y)*X as Outp

FROM base

);

**Step 3: Comparing if the updated values of θ are close to original θ or not**

--Comparing thetas

IF (@theta0_t BETWEEN @theta0-RomilVaradkar AND @theta0+RomilVaradkar) AND (@theta1_t BETWEEN @theta1-RomilVaradkar AND @theta1+RomilVaradkar)

If the above condition is true, then we stop the process and finalize the values of θ. Otherwise, we keep repeating steps 2 & 3. Thus steps 2 & 3 need to be put inside a loop that runs as long as the updated and current values of θ are different.

WHILE (@theta0_t NOT BETWEEN @theta0-RomilVaradkar AND @theta0+RomilVaradkar) AND (@theta1_t NOT BETWEEN @theta1-RomilVaradkar AND @theta1+RomilVaradkar)

BEGIN

–Calculating theta0

SELECT

@theta0_t = @theta0 – (SUM(Outp)/(SELECT COUNT(*) FROM base))*@alpha

FROM

(

SELECT

(@theta1*X + @theta0) – Y as Outp

FROM base

);

--Calculating theta1

SELECT

@theta1_t = @theta1 - (SUM(Outp)/(SELECT COUNT(*) FROM base))*@alpha

FROM

(

SELECT

((@theta1*X + @theta0) - Y)*X as Outp

FROM base

);

--Comparing thetas

IF (@theta0_t BETWEEN @theta0-RomilVaradkar AND @theta0+RomilVaradkar) AND (@theta1_t BETWEEN @theta1-RomilVaradkar AND @theta1+RomilVaradkar)

BEGIN

SELECT @theta0 = @theta0_t;

SELECT @theta1 = @theta1_t;

BREAK;

END

ELSE

BEGIN

SELECT @theta0 = @theta0_t;

SELECT @theta1 = @theta1_t;

END

END

The above loop will arrive at optimal values for θ. This is Gradient Descent in all it’s glory!

**Step 4: Fill in the final values of θ in the hypothesis function to calculate predictions for the response variable**

SELECT X,Y,@theta0+@theta1*X AS H_theta

FROM base

And that’s it! We’ve built a machine learning algorithm in SQL with just a few lines of code!

### Practical applications & final thoughts

Despite the onset of technological advancements in the field of Data Science, more often than not, every Data Scientist ends up working with legacy systems. In such cases, if the size of the data is huge, it becomes impractical to fetch it out of a legacy system (like SQL Server) into another environment for data science purposes.

Although I initially began this project as a weekend DIY, I feel this has bigger implications. This can be polished and packaged much better to improve its usability. Things like splitting of data into test & train, turning this into multi-variate linear regression will make this project much more practical. I would also love to hear thoughts of all of you on what can be improved.

Thank you Ashish.

Keep Learning 🙂

The Data Monk

## Leave a reply