Linear Regression in Snowflake

20 Apr 2021 - 463 Words
Simple Linear Regression in Snowflake

A simple linear regression model is a linear regression model where there is only one independent variable $x$. The dependent variable $y$ is modelled as a linear function of $x$. More simply put the relationship between $x$ and $y$ is described as a straight line with slope $\beta$ (a.k.a gradient) and intercept $\alpha$.

$$ y=\alpha + \beta x $$

In regression we want to find the optimal intercept $\alpha$ and slope $\beta$ to minimise the sum of square errors (or residuals). The residuals are the distances from each point vertically to the model line.

Snowflake ❄️ has some really helpful functions to help with simple linear regression. The three key regression functions are REGR_SLOPE , REGR_INTERCEPT and REGR_R2 used to find the optimal slope, intercept and corresponding r-squared respectively but the rest are useful helpers functions!

All of these functions have the same call signature function_name(y, x) where y is the dependent variable and x is the independent variable in the regression.

The following chart visualises simple linear regression and allows you to vary the slope and intercept and see the impact on the r-squared and other statistics below the chart.

Slope $\beta$ :
Intercept $\alpha$ :
Statistic Formula Value
Sum of square errors (SSE) $\sum (y_i - \hat{y}_i)^2$ 100
Sum of square deviation of model from mean (SSM) $\sum (\hat{y}_i - \bar{y})^2$ 100
r squared (Coefficient of determination) $1 - \frac{SSE}{SST}$ 100
Sample size (number of data points) (REGR_COUNT) $n$ 10
The following stats are for the optimal regression line
Mean of $x$ (REGR_AVGX) $\bar{x} = \frac{1}{n}\sum{x_i}$ 100
Mean of $y$ (REGR_AVGY) $\bar{y} = \frac{1}{n}\sum{y_i}$ 100
$S_{XX}$ (REGR_SXX) $\sum (x_i - \bar{x})^2$ 100
$S_{YY}$ (REGR_SYY) $\sum (y_i - \bar{y})^2$ 100
$S_{XY}$ (REGR_SXY) $\sum (x_i - \bar{x})(y_i - \bar{y})$ 100
Optimal gradient $\beta$ (REGR_SLOPE) $\frac{S_{XY}}{S_{XX}}$ 100
Optimal intercept $\alpha$ (REGR_INTERCEPT) $\bar{y} - \beta\bar{x}$ 100
Optimal r squared - for line with $\alpha$ and $\beta$ (REGR_R2) $\frac{S_{XY}^2}{S_{XX}S_{YY}}$ 100
Correlation $\frac{S_{XY}}{\sqrt{S_{XX}S_{YY}}}$ 100

Before we get started with the sql functions let’s make some toy data to play with! The following query generates the numbers from 1 to 10 for the x column and defines y as (4 * x) + 3 with some noise. We also added a few naughty rows with null values, this will help show off snowflakes functions above.

create temporary table temp_table as (
    select
        row_number() over (order by seq8()) as x,
        4 * x + normal(0, 10, random(10)) + 3 as y
    from table(generator(rowcount => 20)) G

    union all

    select
        *
    from values (null, 100), (50, null)
)

Ok, now that’s done let’s take a look at the table. The data is the same as the chart above.

select
    x,
    y
from temp_table;
X Y
1 -5.940727153
2 -6.435786057
3 12.874649301
4 35.910490613
5 40.969651455
6 27.466977488
7 28.339904759
8 48.022369501
9 36.211253796
10 55.04814023
null 100
50 null

REGR_VALX and REGR_VALY

REGR_VALX, REGR_VALY are helper functions to make sure we only include points (x, y) in the regression where both x and y are not null. REGR_VALX(y , x) returns x if y is not null else it returns null and REGR_VALY(y , x) returns y if x is not null else it returns null.

That means the following two queries are equivalent

select
    REGR_VALX(y , x) as x_val
from temp_table
select
    case when y is not null then x else null end as x_val
from temp_table

Similarly REGR_VALY(y , x) is equivalent to case when x is not null then y else null end

REGR_COUNT

REGR_COUNT is similar to the normal sql count function but only returns the number of non null pairs (x, y). In terms of regression this can be used to calculate the sample size $n$. The following query calculates the same thing 3 times (starting to show how the functions in Snowflake help to simplify)

select
    regr_count(y, x) as n,
    count(regr_valx(y, x)) as n_2,
    count(case when y is not null then x else null end) as n_3
from temp_table;

REGR_AVGX and REGR_AVGY

REGR_AVGX and REGR_AVGY are both similar to the avg function but only include rows where both x and y are non null in the calculation. These are used to calculate the mean values $\bar{x}$ and $\bar{y}$ which are important to calculate the regression gradient, intercept and r-squared. The following query again calculates the same thing three times each for x and y to show how it works.

select
    REGR_AVGX(y, x) as x_mean,
    avg(regr_valx(y, x)) as x_mean_2,
    avg(case when y is not null then x else null end) as x_mean_3,

    REGR_AVGY(y, x) as y_mean,
    avg(regr_valy(y, x)) as y_mean_2,
    avg(case when x is not null then y else null end) as y_mean_3
from temp_table;

REGR_SXX, REGR_SYY and REGR_SXY

SXX, SYY and SXY are useful statistics to calculate the slope, intercept and r-squared. Their formulas are in the table above here.

The following two queries both calculate SXX, SYY and SXY, the first in vanilla SQL the second using the snowflake helper functions.

with averages as (
    select
        avg(case when y is not null then x else null end) as x_mean,
        avg(case when x is not null then y else null end) as y_mean
    from temp_table
)
select
    sum(pow(case when y is not null then x else null end - x_mean, 2)) as sxx,
    sum(pow(case when x is not null then y else null end - y_mean, 2)) as syy,
    sum(
        (case when y is not null then x else null end - x_mean) *
        (case when x is not null then y else null end - y_mean)
    ) as sxy
from temp_table
inner join averages
;

Using the helpers:

select
    regr_sxx(y, x) as sxx,
    regr_syy(y, x) as syy,
    regr_sxy(y, x) as sxy
from temp_table;
;
SXX SYY SXY
82.5 3,992.00 493.48

REGR_SLOPE, REGR_INTERCEPT and REGR_R2

REGR_SLOPE and REGR_INTERCEPT calculate the optimal slope and intercept to minimise the sum of squares of the residuals. The REGR_R2 function calculates the r-squared for the optimal line. This number is between 0 and 1 and explains how much of the variance is explained by the model, the closer the value to 1 the better the fit.

The formulas are given in the tables above here.

Again we show this in vanilla sql and using the snowflake helpers:

with averages as (
    /* First we calculate the average of x and y */
    select
        avg(case when y is not null then x else null end) as x_mean,
        avg(case when x is not null then y else null end) as y_mean
    from temp_table
),

statistics as (
    /* Calculate SXX, SYY and SXY as an intermediate step */
    select
        x_mean,
        y_mean,
        sum(pow(case when y is not null then x else null end - x_mean, 2)) as sxx,
        sum(pow(case when x is not null then y else null end - y_mean, 2)) as syy,
        sum(
            (case when y is not null then x else null end - x_mean) *
            (case when x is not null then y else null end - y_mean)
        ) as sxy
    from temp_table
    inner join averages
    group by 1, 2
)

select
    /* Finally calculate the optimal slope and intercept
     as well as the corresponding r-squared */
    sxy / sxx as slope,
    y_mean - slope * x_mean as intercept,
    (sxy * sxy) / (sxx * syy) as r_squared
from statistics;
;
select
    regr_slope(y, x) as slope,
    regr_intercept(y, x) as intercept,
    regr_r2(y, x) as r_squared
from temp_table;
;
SLOPE INTERCEPT R_SQUARED
5.981534878 -5.651749436 0.7394144386

The functions above allow you to quickly calculate regression values in snowflake and understand the relationships between columns - but remember correlation is not causation s that.

For more about linear regression see my other linear regression article here

Thanks for reading! 👏 Please get in touch with any questions, mistakes or improvements.