Improve model performance with hyperparameter tuning

This tutorial teaches you how to use hyperparameter tuning in BigQuery ML to tune a machine learning model and improve its performance.

You perform hyperparameter tuning by specifying the NUM_TRIALS option of the CREATE MODEL statement, in combination with other model-specific options. When you set these options, BigQuery ML trains multiple versions, or trials of the model, each with slightly different parameters, and returns the trial that performs the best.

This tutorial uses the public tlc_yellow_trips_2018 sample table, which contains information about taxi trip in New York City in 2018.

Create a dataset

Create a BigQuery dataset to store your ML model.

Console

  1. In the Google Cloud console, go to the BigQuery page.

    Go to the BigQuery page

  2. In the Explorer pane, click your project name.

  3. Click View actions > Create dataset

  4. On the Create dataset page, do the following:

    • For Dataset ID, enter bqml_tutorial.

    • For Location type, select Multi-region, and then select US (multiple regions in United States).

    • Leave the remaining default settings as they are, and click Create dataset.

bq

To create a new dataset, use the bq mk command with the --location flag. For a full list of possible parameters, see the bq mk --dataset command reference.

  1. Create a dataset named bqml_tutorial with the data location set to US and a description of BigQuery ML tutorial dataset:

    bq --location=US mk -d \
 --description "BigQuery ML tutorial dataset." \
 bqml_tutorial

    Instead of using the --dataset flag, the command uses the -d shortcut. If you omit -d and --dataset, the command defaults to creating a dataset.

  2. Confirm that the dataset was created:

    bq ls

API

Call the datasets.insert method with a defined dataset resource.

{
  "datasetReference": {
     "datasetId": "bqml_tutorial"
  }
}

BigQuery DataFrames

Before trying this sample, follow the BigQuery DataFrames setup instructions in the BigQuery quickstart using BigQuery DataFrames. For more information, see the BigQuery DataFrames reference documentation.

To authenticate to BigQuery, set up Application Default Credentials. For more information, see Set up ADC for a local development environment. 

import google.cloud.bigquery

bqclient = google.cloud.bigquery.Client()
bqclient.create_dataset("bqml_tutorial", exists_ok=True)

Create a table of training data

Create a table of training data, based on a subset of the tlc_yellow_trips_2018 table data.

Follow these steps to create the table:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    CREATE OR REPLACE TABLE `bqml_tutorial.taxi_tip_input`
AS
SELECT * EXCEPT (tip_amount), tip_amount AS label
FROM
  `bigquery-public-data.new_york_taxi_trips.tlc_yellow_trips_2018`
WHERE
  tip_amount IS NOT NULL
LIMIT 100000;

Create a baseline linear regression model

Create a linear regression model without hyperparameter tuning and train it on the taxi_tip_input table data.

Follow these steps to create the model:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    CREATE OR REPLACE MODEL `bqml_tutorial.baseline_taxi_tip_model`
  OPTIONS (
    MODEL_TYPE = 'LINEAR_REG'
  )
AS
SELECT
  *
FROM
  `bqml_tutorial.taxi_tip_input`;

    The query takes about 2 minutes to complete.

Evaluate the baseline model

Evaluate the performance of the model by using the ML.EVALUATE function. The ML.EVALUATE function evaluates the predicted content ratings returned by the model against the evaluation metrics calculated during model training.

Follow these steps to evaluate the model:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    SELECT *
FROM
  ML.EVALUATE(MODEL `bqml_tutorial.baseline_taxi_tip_model`);

    The results look similar to the following:

    +---------------------+--------------------+------------------------+-----------------------+---------------------+---------------------+
| mean_absolute_error | mean_squared_error | mean_squared_log_error | median_absolute_error |      r2_score       | explained_variance  |
+---------------------+--------------------+------------------------+-----------------------+---------------------+---------------------+
|  2.5853895559690323 | 23760.416358496139 |   0.017392406523370374 | 0.0044248227819481123 | -1934.5450533482465 | -1934.3513857946277 |
+---------------------+--------------------+------------------------+-----------------------+---------------------+---------------------+

The r2_score value for the baseline model is negative, which indicates a poor fit for the data; the closer the R2 score is to 1, the better the model fit is.

Create a linear regression model with hyperparameter tuning

Create a linear regression model with hyperparameter tuning and train it on the taxi_tip_input table data.

You use the following hyperparameter tuning options in the CREATE MODEL statement:

  • The NUM_TRIALS option to set the number of trials to twenty.
  • The MAX_PARALLEL_TRIALS option to run two trials in each training job, for a total of ten jobs and twenty trials. This reduces the training time needed. However, the two concurrent trials don't benefit from each other's training results.
  • The L1_REG option to try different L1 regularization values in the different trials. L1 regularization removes irrelevant features from the model, which helps prevent overfitting.

The other hyperparameter tuning options supported by the model use their default values, as follows:

  • L1_REG: 0
  • HPARAM_TUNING_ALGORITHM: 'VIZIER_DEFAULT'
  • HPARAM_TUNING_OBJECTIVES: ['R2_SCORE']

Follow these steps to create the model:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    CREATE OR REPLACE MODEL `bqml_tutorial.hp_taxi_tip_model`
  OPTIONS (
    MODEL_TYPE = 'LINEAR_REG',
    NUM_TRIALS = 20,
    MAX_PARALLEL_TRIALS = 2,
    L1_REG = HPARAM_RANGE(0, 5))
AS
SELECT
  *
FROM
  `bqml_tutorial.taxi_tip_input`;

    The query takes approximately 20 minutes to complete.

Get information about the training trials

Get information about all of the trials, including their hyperparameter values, objectives, and status, by using the ML.TRIAL_INFO function. This function also returns information about which trial has the best performance, based on this information.

Follow these steps to get trial information:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    SELECT *
FROM
  ML.TRIAL_INFO(MODEL `bqml_tutorial.hp_taxi_tip_model`)
ORDER BY is_optimal DESC;

    The results look similar to the following:

    +----------+-------------------------------------+-----------------------------------+--------------------+--------------------+-----------+---------------+------------+
| trial_id |           hyperparameters           | hparam_tuning_evaluation_metrics  |   training_loss    |     eval_loss      |  status   | error_message | is_optimal |
+----------+-------------------------------------+-----------------------------------+--------------------+--------------------+-----------+---------------+------------+
|        7 |      {"l1_reg":"4.999999999999985"} |  {"r2_score":"0.653653627638174"} | 4.4677841296238165 |  4.478469742512195 | SUCCEEDED | NULL          |       true |
|        2 |  {"l1_reg":"2.402163664510254E-11"} | {"r2_score":"0.6532493667964732"} |  4.457692508421795 |  4.483697081650438 | SUCCEEDED | NULL          |      false |
|        3 |  {"l1_reg":"1.2929452948742316E-7"} |  {"r2_score":"0.653249366811995"} |   4.45769250849513 |  4.483697081449748 | SUCCEEDED | NULL          |      false |
|        4 |  {"l1_reg":"2.5787102060628228E-5"} | {"r2_score":"0.6532493698925899"} |  4.457692523040582 |  4.483697041615808 | SUCCEEDED | NULL          |      false |
|      ... |                             ...     |                           ...     |              ...   |             ...    |       ... |          ...  |        ... |
+----------+-------------------------------------+-----------------------------------+--------------------+--------------------+-----------+---------------+------------+

    The is_optimal column value indicates that trial 7 is the optimal model returned by the tuning.

Evaluate the tuned model trials

Evaluate the performance of the trials by using the ML.EVALUATE function. The ML.EVALUATE function evaluates the predicted content ratings returned by the model against the evaluation metrics calculated during training for all trials.

Follow these steps to evaluate the model trials:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    SELECT *
FROM
  ML.EVALUATE(MODEL `bqml_tutorial.hp_taxi_tip_model`)
ORDER BY r2_score DESC;

    The results look similar to the following:

    +----------+---------------------+--------------------+------------------------+-----------------------+--------------------+--------------------+
| trial_id | mean_absolute_error | mean_squared_error | mean_squared_log_error | median_absolute_error |      r2_score      | explained_variance |
+----------+---------------------+--------------------+------------------------+-----------------------+--------------------+--------------------+
|        7 |   1.151814398002232 |  4.109811493266523 |     0.4918733252641176 |    0.5736103414025084 | 0.6652110305659145 | 0.6652144696114834 |
|       19 |  1.1518143358927102 |  4.109811921460791 |     0.4918672150119582 |    0.5736106106914161 | 0.6652109956848206 | 0.6652144346901685 |
|        8 |   1.152747850702547 |  4.123625876152422 |     0.4897808307399327 |    0.5731702310239184 | 0.6640856984144734 |  0.664088410199906 |
|        5 |   1.152895108945439 |  4.125775524878872 |    0.48939088205957937 |    0.5723300569616766 | 0.6639105860807425 | 0.6639132416838652 |
|      ... |                ...  |                ... |                    ... |                   ... |                ... |                ... |
+----------+---------------------+--------------------+------------------------+-----------------------+--------------------+--------------------+

    The r2_score value for the optimal model, which is trial 7, is 0.66521103056591446, which shows significant improvement over the baseline model.

You can evaluate a specific trial by specifying the TRIAL_ID argument in the ML.EVALUATE function.

For more information about the difference between ML.TRIAL_INFO objectives and ML.EVALUATE evaluation metrics, see Model serving functions.

Use the tuned model to predict taxi tips

Use the optimal model returned by tuning to predict tips for different taxi trips. The optimal model is automatically used by the ML.PREDICT function, unless you select a different trial by specifying the TRIAL_ID argument. The predictions are returned in the predicted_label column.

Follow these steps to get predictions:

  1. In the Google Cloud console, go to the BigQuery page.

    Go to BigQuery

  2. In the query editor, paste in the following query and click Run:

    SELECT *
FROM
  ML.PREDICT(
    MODEL `bqml_tutorial.hp_taxi_tip_model`,
    (
      SELECT
        *
      FROM
        `bqml_tutorial.taxi_tip_input`
      LIMIT 5
    ));

    The results look similar to the following:

    +----------+--------------------+-----------+---------------------+---------------------+-----------------+---------------+-----------+--------------------+--------------+-------------+-------+---------+--------------+---------------+--------------+--------------------+---------------------+----------------+-----------------+-------+
| trial_id |  predicted_label   | vendor_id |   pickup_datetime   |  dropoff_datetime   | passenger_count | trip_distance | rate_code | store_and_fwd_flag | payment_type | fare_amount | extra | mta_tax | tolls_amount | imp_surcharge | total_amount | pickup_location_id | dropoff_location_id | data_file_year | data_file_month | label |
+----------+--------------------+-----------+---------------------+---------------------+-----------------+---------------+-----------+--------------------+--------------+-------------+-------+---------+--------------+---------------+--------------+--------------------+---------------------+----------------+-----------------+-------+
|        7 |  1.343367839584448 | 2         | 2018-01-15 18:55:15 | 2018-01-15 18:56:18 |               1 |             0 | 1         | N                  | 1            |           0 |     0 |       0 |            0 |             0 |            0 | 193                | 193                 |           2018 |               1 |     0 |
|        7 | -1.176072791783461 | 1         | 2018-01-08 10:26:24 | 2018-01-08 10:26:37 |               1 |             0 | 5         | N                  | 3            |        0.01 |     0 |       0 |            0 |           0.3 |         0.31 | 158                | 158                 |           2018 |               1 |     0 |
|        7 |  3.839580104168765 | 1         | 2018-01-22 10:58:02 | 2018-01-22 12:01:11 |               1 |          16.1 | 1         | N                  | 1            |        54.5 |     0 |     0.5 |            0 |           0.3 |         55.3 | 140                | 91                  |           2018 |               1 |     0 |
|        7 |  4.677393985230036 | 1         | 2018-01-16 10:14:35 | 2018-01-16 11:07:28 |               1 |            18 | 1         | N                  | 2            |        54.5 |     0 |     0.5 |            0 |           0.3 |         55.3 | 138                | 67                  |           2018 |               1 |     0 |
|        7 |  7.938988937253062 | 2         | 2018-01-16 07:05:15 | 2018-01-16 08:06:31 |               1 |          17.8 | 1         | N                  | 1            |        54.5 |     0 |     0.5 |            0 |           0.3 |        66.36 | 132                | 255                 |           2018 |               1 | 11.06 |
+----------+--------------------+-----------+---------------------+---------------------+-----------------+---------------+-----------+--------------------+--------------+-------------+-------+---------+--------------+---------------+--------------+--------------------+---------------------+----------------+-----------------+-------+