This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Regression algorithms

Regression is an important and popular machine learning tool that makes predictions from data by learning the relationship between some features of the data and an observed value response.

Regression is an important and popular machine learning tool that makes predictions from data by learning the relationship between some features of the data and an observed value response. Regression is used to make predictions about profits, sales, temperature, stocks, and more. For example, you could use regression to predict the price of a house based on the location, the square footage, the size of the lot, and so on. In this example, the house's value is the response, and the other factors, such as location, are the features.

The optimal set of coefficients found for the regression's equation is known as the model. The relationship between the outcome and the features is summarized in the model, which can then be applied to different data sets, where the outcome value is unknown.

1 - Autoregression

See the autoregressive model example under time series models.

See the autoregressive model example under time series models.

2 - Linear regression

Using linear regression, you can model the linear relationship between independent variables, or features, and a dependent variable, or outcome.

Using linear regression, you can model the linear relationship between independent variables, or features, and a dependent variable, or outcome. You can build linear regression models to:

  • Fit a predictive model to a training data set of independent variables and some dependent variable. Doing so allows you to use feature variable values to make predictions on outcomes. For example, you can predict the amount of rain that will fall on a particular day of the year.

  • Determine the strength of the relationship between an independent variable and some outcome variable. For example, suppose you want to determine the importance of various weather variables on the outcome of how much rain will fall. You can build a linear regression model based on observations of weather patterns and rainfall to find the answer.

Unlike Logistic regression, which you use to determine a binary classification outcome, linear regression is primarily used to predict continuous numerical outcomes in linear relationships.

You can use the following functions to build a linear regression model, view the model, and use the model to make predictions on a set of test data:

For a complete example of how to use linear regression on a table in Vertica, see Building a linear regression model.

2.1 - Building a linear regression model

This linear regression example uses a small data set named faithful.

This linear regression example uses a small data set named faithful. The data set gives the waiting time between eruptions and the duration of the eruption for the Old Faithful geyser in Yellowstone National Park. The duration of each eruption can last anywhere between 1.5 and 5 minutes. The time between eruptions and the length of each eruption varies. However, you can estimate the time of the next eruption based on the duration of the previous eruption. The example shows how you can build a model to predict the value of eruptions, given the value of the waiting feature.

Before you begin the example, load the Machine Learning sample data.

  1. Create the linear regression model, named linear_reg_faithful, using the faithful_training training data:

    => SELECT LINEAR_REG('linear_reg_faithful', 'faithful_training', 'eruptions', 'waiting'
                      USING PARAMETERS optimizer='BFGS');
        LINEAR_REG
---------------------------
 Finished in 6 iterations

(1 row)

  2. View the summary output of linear_reg_faithful:

    => SELECT GET_MODEL_SUMMARY(USING PARAMETERS model_name='linear_reg_faithful');
--------------------------------------------------------------------------------
=======
details
=======
predictor|coefficient|std_err |t_value |p_value
---------+-----------+--------+--------+--------
Intercept| -2.06795  | 0.21063|-9.81782| 0.00000
waiting  |  0.07876  | 0.00292|26.96925| 0.00000

==============
regularization
==============
type| lambda
----+--------
none| 1.00000

===========
call_string
===========
linear_reg('public.linear_reg_faithful', 'faithful_training', '"eruptions"', 'waiting'
USING PARAMETERS optimizer='bfgs', epsilon=1e-06, max_iterations=100,
regularization='none', lambda=1)

===============
Additional Info
===============
Name              |Value
------------------+-----
iteration_count   |  3
rejected_row_count|  0
accepted_row_count| 162
(1 row)

  3. Create a new table that contains the response values from running the PREDICT_LINEAR_REG function on your test data. Name this table pred_faithful_results. View the results in the pred_faithful_results table:

    => CREATE TABLE pred_faithful_results AS
       (SELECT id, eruptions, PREDICT_LINEAR_REG(waiting USING PARAMETERS model_name='linear_reg_faithful')
        AS pred FROM faithful_testing);
CREATE TABLE

=> SELECT * FROM pred_faithful_results ORDER BY id;
 id  | eruptions |       pred
-----+-----------+------------------
   4 |     2.283 |  2.8151271587036
   5 |     4.533 | 4.62659045686076
   8 |       3.6 | 4.62659045686076
   9 |      1.95 | 1.94877514654148
  11 |     1.833 | 2.18505296804024
  12 |     3.917 | 4.54783118302784
  14 |      1.75 |  1.6337380512098
  20 |      4.25 | 4.15403481386324
  22 |      1.75 |  1.6337380512098
.
.
.
(110 rows)

Calculating the mean squared error (MSE)

You can calculate how well your model fits the data is by using the MSE function. MSE returns the average of the squared differences between actual value and predicted values.

=> SELECT MSE (eruptions::float, pred::float) OVER() FROM
(SELECT eruptions, pred FROM pred_faithful_results) AS prediction_output;
        mse        |                   Comments
-------------------+-----------------------------------------------
 0.252925741352641 | Of 110 rows, 110 were used and 0 were ignored
(1 row)

See also

3 - Random forest for regression

The Random Forest for regression algorithm creates an ensemble model of regression trees.

The Random Forest for regression algorithm creates an ensemble model of regression trees. Each tree is trained on a randomly selected subset of the training data. The algorithm predicts the value that is the mean prediction of the individual trees.

You can use the following functions to train the Random Forest model, and use the model to make predictions on a set of test data:

For a complete example of how to use the Random Forest for regression algorithm in Vertica, see Building a random forest regression model.

3.1 - Building a random forest regression model

This example uses the "mtcars" dataset to create a random forest model to predict the value of carb (the number of carburetors).

This example uses the "mtcars" dataset to create a random forest model to predict the value of carb (the number of carburetors).

Before you begin the example, load the Machine Learning sample data.

  1. Use RF_REGRESSOR to create the random forest model myRFRegressorModel using the mtcars training data. View the summary output of the model with GET_MODEL_SUMMARY:

    => SELECT RF_REGRESSOR ('myRFRegressorModel', 'mtcars', 'carb', 'mpg, cyl, hp, drat, wt' USING PARAMETERS
ntree=100, sampling_size=0.3);
RF_REGRESSOR
--------------
Finished
(1 row)


=> SELECT GET_MODEL_SUMMARY(USING PARAMETERS model_name='myRFRegressorModel');
--------------------------------------------------------------------------------
===========
call_string
===========
SELECT rf_regressor('public.myRFRegressorModel', 'mtcars', '"carb"', 'mpg, cyl, hp, drat, wt'
USING PARAMETERS exclude_columns='', ntree=100, mtry=1, sampling_size=0.3, max_depth=5, max_breadth=32,
min_leaf_size=5, min_info_gain=0, nbins=32);

=======
details
=======
predictor|type
---------+-----
mpg      |float
cyl      | int
hp       | int
drat     |float
wt       |float
===============
Additional Info
===============
Name              |Value
------------------+-----
tree_count        | 100
rejected_row_count|  0
accepted_row_count| 32
(1 row)

  2. Use PREDICT_RF_REGRESSOR to predict the number of carburetors:

    => SELECT PREDICT_RF_REGRESSOR (mpg,cyl,hp,drat,wt
USING PARAMETERS model_name='myRFRegressorModel') FROM mtcars;
PREDICT_RF_REGRESSOR
----------------------
2.94774203574204
2.6954087024087
2.6954087024087
2.89906346431346
2.97688489288489
2.97688489288489
2.7086587024087
2.92078965478965
2.97688489288489
2.7086587024087
2.95621822621823
2.82255155955156
2.7086587024087
2.7086587024087
2.85650394050394
2.85650394050394
2.97688489288489
2.95621822621823
2.6954087024087
2.6954087024087
2.84493251193251
2.97688489288489
2.97688489288489
2.8856467976468
2.6954087024087
2.92078965478965
2.97688489288489
2.97688489288489
2.7934087024087
2.7934087024087
2.7086587024087
2.72469441669442
(32 rows)

4 - SVM (support vector machine) for regression

Support Vector Machine (SVM) for regression predicts continuous ordered variables based on the training data.

Support Vector Machine (SVM) for regression predicts continuous ordered variables based on the training data.

Unlike Logistic regression, which you use to determine a binary classification outcome, SVM for regression is primarily used to predict continuous numerical outcomes.

You can use the following functions to build an SVM for regression model, view the model, and use the model to make predictions on a set of test data:

For a complete example of how to use the SVM algorithm in Vertica, see Building an SVM for regression model.

4.1 - Building an SVM for regression model

This SVM for regression example uses a small data set named faithful, based on the Old Faithful geyser in Yellowstone National Park.

This SVM for regression example uses a small data set named faithful, based on the Old Faithful geyser in Yellowstone National Park. The data set contains values about the waiting time between eruptions and the duration of eruptions of the geyser. The example shows how you can build a model to predict the value of eruptions, given the value of the waiting feature.

Before you begin the example, load the Machine Learning sample data.

  1. Create the SVM model, named svm_faithful, using the faithful_training training data:

    => SELECT SVM_REGRESSOR('svm_faithful', 'faithful_training', 'eruptions', 'waiting'
                      USING PARAMETERS error_tolerance=0.1, max_iterations=100);
        SVM_REGRESSOR
---------------------------
 Finished in 5 iterations
Accepted Rows: 162   Rejected Rows: 0
(1 row)

  2. View the summary output of svm_faithful:

    => SELECT GET_MODEL_SUMMARY(USING PARAMETERS model_name='svm_faithful');

------------------------------------------------------------------
=======
details
=======

===========================
Predictors and Coefficients
===========================
         |Coefficients
---------+------------
Intercept|  -1.59007
waiting  |   0.07217
===========
call_string
===========
Call string:
SELECT svm_regressor('public.svm_faithful', 'faithful_training', '"eruptions"',
'waiting'USING PARAMETERS error_tolerance = 0.1, C=1, max_iterations=100,
epsilon=0.001);

===============
Additional Info
===============
Name              |Value
------------------+-----
accepted_row_count| 162
rejected_row_count|  0
iteration_count  |  5
(1 row)

  3. Create a new table that contains the response values from running the PREDICT_SVM_REGRESSOR function on your test data. Name this table pred_faithful_results. View the results in the pred_faithful_results table:

    => CREATE TABLE pred_faithful AS
       (SELECT id, eruptions, PREDICT_SVM_REGRESSOR(waiting USING PARAMETERS model_name='svm_faithful')
        AS pred FROM faithful_testing);
CREATE TABLE
=> SELECT * FROM pred_faithful ORDER BY id;
 id  | eruptions |       pred
-----+-----------+------------------
   4 |     2.283 | 2.88444568755189
   5 |     4.533 | 4.54434581879796
   8 |       3.6 | 4.54434581879796
   9 |      1.95 | 2.09058040739072
  11 |     1.833 | 2.30708912016195
  12 |     3.917 | 4.47217624787422
  14 |      1.75 | 1.80190212369576
  20 |      4.25 | 4.11132839325551
  22 |      1.75 | 1.80190212369576
.
.
.
(110 rows)

Calculating the mean squared error (MSE)

You can calculate how well your model fits the data is by using the MSE function. MSE returns the average of the squared differences between actual value and predicted values.

=> SELECT MSE(obs::float, prediction::float) OVER()
   FROM (SELECT eruptions AS obs, pred AS prediction
         FROM pred_faithful) AS prediction_output;
        mse        |                   Comments
-------------------+-----------------------------------------------
 0.254499811834235 | Of 110 rows, 110 were used and 0 were ignored
(1 row)

See also

5 - XGBoost for regression

The following XGBoost functions create and perform predictions with a regression model:.

XGBoost (eXtreme Gradient Boosting) is a popular supervised-learning algorithm used for regression and classification on large datasets. It uses sequentially-built shallow decision trees to provide accurate results and a highly-scalable training method that avoids overfitting.

The following XGBoost functions create and perform predictions with a regression model:

Example

This example uses a small data set named "mtcars", which contains design and performance data for 32 automobiles from 1973-1974, and creates an XGBoost regression model to predict the value of the variable carb (the number of carburetors).

Before you begin the example, load the Machine Learning sample data.

  1. Use XGB_REGRESSOR to create the XGBoost regression model xgb_cars from the mtcars dataset:

    => SELECT XGB_REGRESSOR ('xgb_cars', 'mtcars', 'carb', 'mpg, cyl, hp, drat, wt'
    USING PARAMETERS learning_rate=0.5);
 XGB_REGRESSOR
---------------
 Finished
(1 row)

    You can then view a summary of the model with GET_MODEL_SUMMARY:

    
=> SELECT GET_MODEL_SUMMARY(USING PARAMETERS model_name='xgb_cars');
                  GET_MODEL_SUMMARY
------------------------------------------------------
===========
call_string
===========
xgb_regressor('public.xgb_cars', 'mtcars', '"carb"', 'mpg, cyl, hp, drat, wt'
USING PARAMETERS exclude_columns='', max_ntree=10, max_depth=5, nbins=32, objective=squarederror,
split_proposal_method=global, epsilon=0.001, learning_rate=0.5, min_split_loss=0, weight_reg=0, sampling_size=1)

=======
details
=======
predictor|      type
---------+----------------
   mpg   |float or numeric
   cyl   |      int
   hp    |      int
  drat   |float or numeric
   wt    |float or numeric

===============
Additional Info
===============
       Name       |Value
------------------+-----
    tree_count    | 10
rejected_row_count|  0
accepted_row_count| 32

(1 row)

  2. Use PREDICT_XGB_REGRESSOR to predict the number of carburetors:

    => SELECT carb, PREDICT_XGB_REGRESSOR (mpg,cyl,hp,drat,wt USING PARAMETERS model_name='xgb_cars') FROM mtcars;
 carb | PREDICT_XGB_REGRESSOR
------+-----------------------
    4 |      4.00335213618023
    2 |       2.0038188946536
    6 |      5.98866003194438
    1 |      1.01774386191546
    2 |       1.9959801016274
    2 |       2.0038188946536
    4 |      3.99545403625739
    8 |      7.99211056556231
    2 |      1.99291901733151
    3 |       2.9975688946536
    3 |       2.9975688946536
    1 |      1.00320357711227
    2 |       2.0038188946536
    4 |      3.99545403625739
    4 |      4.00124134679445
    1 |      1.00759516721382
    4 |      3.99700517763435
    4 |      3.99580193056138
    4 |      4.00009088187525
    3 |       2.9975688946536
    2 |      1.98625064560888
    1 |      1.00355294416998
    2 |      2.00666247039502
    1 |      1.01682931210169
    4 |      4.00124134679445
    1 |      1.01007809485918
    2 |      1.98438405824605
    4 |      3.99580193056138
    2 |      1.99291901733151
    4 |      4.00009088187525
    2 |       2.0038188946536
    1 |      1.00759516721382
(32 rows)