Feature Engineering in MindsDB¶

The more data you have, the more accurate predictions you get.

We recommend you provide the predictor with as many historical data rows and data columns as possible to make your predictions even more accurate. The examples presented here prove this hypothesis.

If you want to follow the examples, please sign up for the MindsDB Cloud account here.

Prerequisites¶

The base table is available in the example_db integration in the MindsDB Cloud Editor. In order to be able to use it, you must first create a database like this:

CREATE DATABASE example_db
WITH ENGINE = "postgres",
PARAMETERS = {
    "user": "demo_user",
    "password": "demo_password",
    "host": "3.220.66.106",
    "port": "5432",
    "database": "demo"
};

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Once that's done, you can run the following commands with us.

Example: Adding More Data Columns¶

Introduction¶

Here, we'll create several predictors using the same table, increasing the number of data columns with each step.

We start with the base table and create a predictor based on it. Then we add two columns to our base table and again create a predictor based on the enhanced table. At last, we add another two columns and create a predictor.

By comparing the accuracies of the predictors, we'll find that more data results in more accurate predictions.

Let's get started.

Let's Run the Codes¶

Here, we go through the codes for the base table and enhanced base tables simultaneously.

Data Setup¶

Let's prepare and verify the data. Here, we create the views and query them to ensure the input for the predictors is in order.

Using the Base TableUsing the Base Table + 2 More ColumnsUsing the Base Table + 4 More Columns

Let's start by querying the data from the example_db.demo_data.used_car_price table, which is our base table.

SELECT *
FROM example_db.demo_data.used_car_price
LIMIT 5;

On execution, we get:

+-----+----+-----+------------+-------+--------+---+----+----------+
|model|year|price|transmission|mileage|fueltype|tax|mpg |enginesize|
+-----+----+-----+------------+-------+--------+---+----+----------+
| A1  |2017|12500|Manual      |15735  |Petrol  |150|55.4|1.4       |
| A6  |2016|16500|Automatic   |36203  |Diesel  |20 |64.2|2         |
| A1  |2016|11000|Manual      |29946  |Petrol  |30 |55.4|1.4       |
| A4  |2017|16800|Automatic   |25952  |Diesel  |145|67.3|2         |
| A3  |2019|17300|Manual      |1998   |Petrol  |145|49.6|1         |
+-----+----+-----+------------+-------+--------+---+----+----------+

Where:

Name	Description
`model`	Model of the car.
`year`	Year of production.
`price`	Price of the car.
`transmission`	Transmission (`Manual`, or `Automatic`, or `Semi-Auto`).
`mileage`	Mileage of the car.
`fueltype`	Fuel type of the car.
`tax`	Tax.
`mpg`	Miles per gallon.
`enginesize`	Engine size of the car.

Let's create a view based on the example_db.demo_data.used_car_price table, and add two more columns. Please note that we replace the mpg column with the kml column.

The added columns are:
- the kml column, calculated from the mpg column using the formula like in the query below, stands for kilometers per liter,
- the years_old column, calculated by subtracting car's year from the current date, stands for car's age.

CREATE VIEW used_car_price_plus_2_columns (
SELECT * FROM example_db (
    SELECT 
    model, 
    year, 
    price, 
    transmission, 
    mileage, 
    fueltype, 
    tax,
    enginesize, 
    ROUND(CAST((mpg / 2.3521458) AS numeric), 1) AS kml, -- mpg (miles per galon) is replaced with kml (kilometers per liter)
    (date_part('year', CURRENT_DATE)-year) AS years_old -- age of a car
    FROM demo_data.used_car_price
)
);

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Let's query the newly created view.

SELECT *
FROM views.used_car_price_plus_2_columns
LIMIT 5;

On execution, we get:

+-----+----+-----+------------+-------+--------+---+----+----------+----+---------+
|model|year|price|transmission|mileage|fueltype|tax|mpg |enginesize|kml |years_old|
+-----+----+-----+------------+-------+--------+---+----+----------+----+---------+
| A1  |2017|12500|Manual      |15735  |Petrol  |150|55.4|1.4       |23.6|5        |
| A6  |2016|16500|Automatic   |36203  |Diesel  |20 |64.2|2         |27.3|6        |
| A1  |2016|11000|Manual      |29946  |Petrol  |30 |55.4|1.4       |23.6|6        |
| A4  |2017|16800|Automatic   |25952  |Diesel  |145|67.3|2         |28.6|5        |
| A3  |2019|17300|Manual      |1998   |Petrol  |145|49.6|1         |21.1|3        |
+-----+----+-----+------------+-------+--------+---+----+----------+----+---------+

Let's create a view based on the example_db.demo_data.used_car_price table, and add four more columns. Please note that we replace the mpg column with the kml column.

The added columns are:
- the kml column, calculated from the mpg column using the formula like in the query below, stands for kilometers per liter,
- the years_old column, calculated by subtracting car's year from the current date, stands for car's age,
- the units_to_sell column, calculated using the OVER and PARTITION BY clauses, indicates how many units of a certain car model are sold in a year,
- the tax_div_price column, calculated by dividing the tax column by the price column.

CREATE VIEW used_car_price_plus_another_2_columns (
SELECT * FROM example_db (
    SELECT 
    model, 
    year, 
    price, 
    transmission, 
    mileage, 
    fueltype, 
    tax,
    enginesize, 
    ROUND(CAST((mpg / 2.3521458) AS numeric), 1) AS kml, -- mpg (miles per galon) is replaced with kml (kilometers per liter)
    (date_part('year', CURRENT_DATE)-year) AS years_old, -- age of a car
    COUNT(*) OVER (PARTITION BY model, year) AS units_to_sell, -- how many units of a certain model are sold in a year
    ROUND((CAST(tax AS decimal) / price), 3) AS tax_div_price -- value of tax divided by price of a car
    FROM demo_data.used_car_price
)
);

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Let's query the newly created view.

SELECT *
FROM views.used_car_price_plus_another_2_columns
LIMIT 5;

On execution, we get:

+-----+----+-----+------------+-------+--------+---+----+----------+----+---------+-------------+-------------+
|model|year|price|transmission|mileage|fueltype|tax|mpg |enginesize|kml |years_old|units_to_sell|tax_div_price|
+-----+----+-----+------------+-------+--------+---+----+----------+----+---------+-------------+-------------+
| A1  |2010|9990 |Automatic   |38000  |Petrol  |125|53.3|1.4       |22.7|12       |1            |0.013        |
| A1  |2011|6995 |Manual      |65000  |Petrol  |125|53.3|1.4       |22.7|11       |5            |0.018        |
| A1  |2011|6295 |Manual      |107000 |Petrol  |125|53.3|1.4       |22.7|11       |5            |0.020        |
| A1  |2011|4250 |Manual      |116000 |Diesel  |20 |70.6|1.6       |30.0|11       |5            |0.005        |
| A1  |2011|6475 |Manual      |45000  |Diesel  |0  |70.6|1.6       |30.0|11       |5            |0.000        |
+-----+----+-----+------------+-------+--------+---+----+----------+----+---------+-------------+-------------+

Dropping a View

If you want to drop a view, run the command DROP VIEW view_name;.

Creating Predictors¶

Now, we create predictors based on the example_db.demo_data.used_car_price table and its extensions.

Using the Base TableUsing the Base Table + 2 More ColumnsUsing the Base Table + 4 More Columns

CREATE PREDICTOR mindsdb.price_predictor
FROM example_db
(SELECT * FROM demo_data.used_car_price)
PREDICT price;

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

CREATE PREDICTOR mindsdb.price_predictor_plus_2_columns
FROM views
(SELECT * FROM used_car_price_plus_2_columns)
PREDICT price;

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

CREATE PREDICTOR mindsdb.price_predictor_plus_another_2_columns
FROM views
(SELECT * FROM used_car_price_plus_another_2_columns)
PREDICT price;

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Dropping a Predictor

If you want to drop a predictor, run the command DROP PREDICTOR predictor_name;.

Predictor Status¶

Finally, let's check the predictor status whose value is generating at first, then training, and at last, complete.

Using the Base TableUsing the Base Table + 2 More ColumnsUsing the Base Table + 4 More Columns

SELECT *
FROM mindsdb.predictors
WHERE name='price_predictor';

On execution, we get:

+---------------+--------+--------+---------+-------------+---------------+------+--------------------------------------+----------------+
|name           |status  |accuracy|predict  |update_status|mindsdb_version|error |select_data_query                     |training_options|
+---------------+--------+--------+---------+-------------+---------------+------+--------------------------------------+----------------+
|price_predictor|complete|0.963   |price    |up_to_date   |22.10.2.1      |[NULL]|SELECT * FROM demo_data.used_car_price|                |
+---------------+--------+--------+---------+-------------+---------------+------+--------------------------------------+----------------+

SELECT *
FROM mindsdb.predictors
WHERE name='price_predictor_plus_2_columns';

On execution, we get:

+------------------------------+--------+--------+---------+-------------+---------------+------+-------------------------------------------+----------------+
|name                          |status  |accuracy|predict  |update_status|mindsdb_version|error |select_data_query                          |training_options|
+------------------------------+--------+--------+---------+-------------+---------------+------+-------------------------------------------+----------------+
|price_predictor_plus_2_columns|complete|0.965   |price    |up_to_date   |22.10.2.1      |[NULL]|SELECT * FROM used_car_price_plus_2_columns|                |
+------------------------------+--------+--------+---------+-------------+---------------+------+-------------------------------------------+----------------+

SELECT *
FROM mindsdb.predictors
WHERE name='price_predictor_plus_another_2_columns';

On execution, we get:

+--------------------------------------+--------+--------+---------+-------------+---------------+------+---------------------------------------------------+----------------+
|name                                  |status  |accuracy|predict  |update_status|mindsdb_version|error |select_data_query                                  |training_options|
+--------------------------------------+--------+--------+---------+-------------+---------------+------+---------------------------------------------------+----------------+
|price_predictor_plus_another_2_columns|complete|0.982   |price    |up_to_date   |22.10.2.1      |[NULL]|SELECT * FROM used_car_price_plus_another_2_columns|                |
+--------------------------------------+--------+--------+---------+-------------+---------------+------+---------------------------------------------------+----------------+

Accuracy Comparison¶

Once the training process of all three predictors completes, we see the accuracy values.

For the base table, we get an accuracy value of 0.963.
For the base table with two more data columns, we get an accuracy value of 0.965. The accuracy value increased, as expected.
For the base table with four more data columns, we get an accuracy value of 0.982. The accuracy value increased again, as expected.

True vs Predicted Price Comparison¶

Let's compare how close the predicted price values are to the true price.

+-------+-------+---------------+-----------+-----------+--------------+----------------+----------------+---------------+
| model | year  | transmission  | fueltype  | mileage   | true_price   | pred_price_1   | pred_price_2   | pred_price_3  |
+-------+-------+---------------+-----------+-----------+--------------+----------------+----------------+---------------+
| A1    | 2017  | Manual        | Petrol    | 7620      | 14440        | 17268          | 17020          | 14278         |
| A6    | 2016  | Automatic     | Diesel    | 20335     | 18982        | 17226          | 17935          | 19016         |
| A3    | 2018  | Semi-Auto     | Diesel    | 9058      | 19900        | 25641          | 23008          | 21286         |
+-------+-------+---------------+-----------+-----------+--------------+----------------+----------------+---------------+

The prices predicted by the third predictor, having the highest accuracy value, are the closest to the true price, as expected.

Example: Joining Data Tables¶

Introduction¶

We start by creating a predictor from the car_sales table. Then, we add more data by joining the car_sales and car_info tables. We create a predictor based on the car_sales_info view.

Let's get started.

Let's Run the Codes¶

Here, we go through the codes using partial tables and the full table after joining the data.

Data Setup¶

Here is the car_sales table:

SELECT *
FROM example_db.demo_data.car_sales
LIMIT 5;

On execution, we get:

+-----+----+-----+------------+-------+--------+---+
|model|year|price|transmission|mileage|fueltype|tax|
+-----+----+-----+------------+-------+--------+---+
| A1  |2017|12500|Manual      |15735  |Petrol  |150|
| A6  |2016|16500|Automatic   |36203  |Diesel  |20 |
| A1  |2016|11000|Manual      |29946  |Petrol  |30 |
| A4  |2017|16800|Automatic   |25952  |Diesel  |145|
| A3  |2019|17300|Manual      |1998   |Petrol  |145|
+-----+----+-----+------------+-------+--------+---+

Where:

Name	Description
`model`	Model of the car.
`year`	Year of production.
`price`	Price of the car.
`transmission`	Transmission (`Manual`, or `Automatic`, or `Semi-Auto`).
`mileage`	Mileage of the car.
`fueltype`	Fuel type of the car.
`tax`	Tax.

And here is the car_info table:

SELECT *
FROM example_db.demo_data.car_info
LIMIT 5;

On execution, we get:

+-----+----+------------+---------+-----+----------+
|model|year|transmission|fueltype |mpg  |enginesize|
+-----+----+------------+---------+-----+----------+
| A1  |2010|Automatic   |Petrol   |53.3 |1.4       |
| A1  |2011|Manual      |Diesel   |70.6 |1.6       |
| A1  |2011|Manual      |Petrol   |53.3 |1.4       |
| A1  |2012|Automatic   |Petrol   |50.6 |1.4       |
| A1  |2012|Manual      |Diesel   |72.95|1.7       |
+-----+----+------------+---------+-----+----------+

Where:

Name	Description
`model`	Model of the car.
`year`	Year of production.
`transmission`	Transmission (`Manual`, or `Automatic`, or `Semi-Auto`).
`fueltype`	Fuel type of the car.
`mpg`	Miles per gallon.
`enginesize`	Engine size of the car.

Let's join the car_sales and car_info tables on the model, year, transmission, and fueltype columns.

SELECT * FROM example_db
(
  SELECT s.*, i.mpg, i.enginesize
  FROM demo_data.car_sales s
  JOIN demo_data.car_info i
  ON s.model=i.model
  AND s.year=i.year
  AND s.transmission=i.transmission
  AND s.fueltype=i.fueltype
)
LIMIT 5;

Nested SELECT Statements

Please note that we use the nested SELECT statement in order to trigger native query at the MindsDB Cloud Editor. Here, the example_db database is a PostgreSQL database, so we trigger PostgreSQL-native syntax.

On execution, we get:

+-----+----+-----+------------+-------+--------+---+----+----------+
|model|year|price|transmission|mileage|fueltype|tax|mpg |enginesize|
+-----+----+-----+------------+-------+--------+---+----+----------+
| A1  |2010|9990 |Automatic   |38000  |Petrol  |125|53.3|1.4       |
| A1  |2011|4250 |Manual      |116000 |Diesel  |20 |70.6|1.6       |
| A1  |2011|6475 |Manual      |45000  |Diesel  |0  |70.6|1.6       |
| A1  |2011|6295 |Manual      |107000 |Petrol  |125|53.3|1.4       |
| A1  |2011|7495 |Manual      |60700  |Petrol  |125|53.3|1.4       |
+-----+----+-----+------------+-------+--------+---+----+----------+

Now, we create a view based on the JOIN query:

CREATE VIEW car_sales_info 
(
    SELECT * FROM example_db
    (
    SELECT s.*, i.mpg, i.enginesize
    FROM demo_data.car_sales s
    JOIN demo_data.car_info i
    ON s.model=i.model
    AND s.year=i.year
    AND s.transmission=i.transmission
    AND s.fueltype=i.fueltype
    )
);

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Let's verify the view by selecting from it.

SELECT *
FROM views.car_sales_info
LIMIT 5;

On execution, we get:

+-----+----+-----+------------+-------+--------+---+----+----------+
|model|year|price|transmission|mileage|fueltype|tax|mpg |enginesize|
+-----+----+-----+------------+-------+--------+---+----+----------+
| A1  |2010|9990 |Automatic   |38000  |Petrol  |125|53.3|1.4       |
| A1  |2011|4250 |Manual      |116000 |Diesel  |20 |70.6|1.6       |
| A1  |2011|6475 |Manual      |45000  |Diesel  |0  |70.6|1.6       |
| A1  |2011|6295 |Manual      |107000 |Petrol  |125|53.3|1.4       |
| A1  |2011|7495 |Manual      |60700  |Petrol  |125|53.3|1.4       |
+-----+----+-----+------------+-------+--------+---+----+----------+

Creating Predictors¶

Let's create a predictor with the car_sales table as input data.

CREATE PREDICTOR mindsdb.price_predictor_car_sales
FROM example_db
  (SELECT * FROM demo_data.car_sales)
PREDICT price;

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Now, let's create a predictor for the table that is a JOIN between the car_sales and car_info tables.

CREATE PREDICTOR mindsdb.price_predictor_car_sales_info
FROM views
  (SELECT * FROM car_sales_info)
PREDICT price;

On execution, we get:

Query OK, 0 rows affected (x.xxx sec)

Predictor Status¶

Next, we check the status of both predictors.

We start with the predictor based on the partial table.

SELECT *
FROM mindsdb.predictors
WHERE name='price_predictor_car_sales';

On execution, we get:

+-------------------------+--------+--------+---------+-------------+---------------+------+---------------------------------+----------------+
|name                     |status  |accuracy|predict  |update_status|mindsdb_version|error |select_data_query                |training_options|
+-------------------------+--------+--------+---------+-------------+---------------+------+---------------------------------+----------------+
|price_predictor_car_sales|complete|0.912   |price    |up_to_date   |22.10.2.1      |[NULL]|SELECT * FROM demo_data.car_sales|                |
+-------------------------+--------+--------+---------+-------------+---------------+------+---------------------------------+----------------+

And now, for the predictor based on the full table.

SELECT *
FROM mindsdb.predictors
WHERE name='price_predictor_car_sales_info';

On execution, we get:

+------------------------------+--------+--------+---------+-------------+---------------+------+----------------------------+----------------+
|name                          |status  |accuracy|predict  |update_status|mindsdb_version|error |select_data_query           |training_options|
+------------------------------+--------+--------+---------+-------------+---------------+------+----------------------------+----------------+
|price_predictor_car_sales_info|complete|0.912   |price    |up_to_date   |22.10.2.1      |[NULL]|SELECT * FROM car_sales_info|                |
+------------------------------+--------+--------+---------+-------------+---------------+------+----------------------------+----------------+

Accuracy Comparison¶

The accuracy values are 0.912 for both the predictors. The predictor already learns how the combination of model+year+transmission+fueltype affects the price, so joining more data columns doesn't play a role in this particular example.