The CREATE MODEL statement for matrix factorization models

This document describes the CREATE MODEL statement for creating matrix factorization models in BigQuery. Matrix factorization models use collaborative filtering to identify similar items. Use matrix factorization models with the ML.RECOMMEND function to generate recommendations.

Matrix factorization models support hyperparameter tuning.

For information about the supported SQL statements and functions for each model type, see End-to-end user journey for each model.

CREATE MODEL syntax

{CREATE MODEL | CREATE MODEL IF NOT EXISTS | CREATE OR REPLACE MODEL} model_name
OPTIONS(model_option_list)
AS query_statement

model_option_list:
MODEL_TYPE = 'MATRIX_FACTORIZATION'
    [, FEEDBACK_TYPE = {'EXPLICIT' | 'IMPLICIT'} ]
    [, NUM_FACTORS = { int64_value | HPARAM_RANGE(range) | HPARAM_CANDIDATES([candidates]) } ]
    [, USER_COL = string_value ]
    [, ITEM_COL = string_value ]
    [, RATING_COL = string_value ]
    [, WALS_ALPHA = { float64_value | HPARAM_RANGE(range) | HPARAM_CANDIDATES([candidates]) } ]
    [, L2_REG = { float64_value | HPARAM_RANGE(range) | HPARAM_CANDIDATES([candidates]) } ]
    [, MAX_ITERATIONS = int64_value ]
    [, EARLY_STOP = { TRUE | FALSE } ]
    [, MIN_REL_PROGRESS = float64_value ]
    [, DATA_SPLIT_METHOD = { 'AUTO_SPLIT' | 'RANDOM' | 'CUSTOM' | 'SEQ' | 'NO_SPLIT' } ]
    [, DATA_SPLIT_EVAL_FRACTION = float64_value ]
    [, DATA_SPLIT_TEST_FRACTION = float64_value ]
    [, DATA_SPLIT_COL = string_value ]
    [, NUM_TRIALS = int64_value ]
    [, MAX_PARALLEL_TRIALS = int64_value ]
    [, HPARAM_TUNING_ALGORITHM = { 'VIZIER_DEFAULT' | 'RANDOM_SEARCH' | 'GRID_SEARCH' } ]
    [, HPARAM_TUNING_OBJECTIVES = { 'MEAN_SQUARED_ERROR' | 'MEAN_AVERAGE_PRECISION' | ... } ]
    [, MODEL_REGISTRY = { 'VERTEX_AI' } ]
    [, VERTEX_AI_MODEL_ID = string_value ]
    [, VERTEX_AI_MODEL_VERSION_ALIASES = string_array ]
    [, KMS_KEY_NAME = string_value ]

CREATE MODEL

Creates and trains a new model in the specified dataset. If the model name exists, CREATE MODEL returns an error.

CREATE MODEL IF NOT EXISTS

Creates and trains a new model only if the model doesn't exist in the specified dataset.

CREATE OR REPLACE MODEL

Creates and trains a model and replaces an existing model with the same name in the specified dataset.

model_name

The name of the model you're creating or replacing. The model name must be unique in the dataset: no other model or table can have the same name. The model name must follow the same naming rules as a BigQuery table. A model name can:

  • Contain up to 1,024 characters
  • Contain letters (upper or lower case), numbers, and underscores

model_name is not case-sensitive.

If you don't have a default project configured, then you must prepend the project ID to the model name in the following format, including backticks:

`[PROJECT_ID].[DATASET].[MODEL]`

For example, `myproject.mydataset.mymodel`.

MODEL_TYPE

Syntax

MODEL_TYPE = 'MATRIX_FACTORIZATION'

Description

Specifies the model type. To create a matrix factorization model, specify MATRIX_FACTORIZATION.

FEEDBACK_TYPE

Syntax

FEEDBACK_TYPE = { 'EXPLICIT' | 'IMPLICIT' }

Description

Specifies the feedback type for the model. The feedback type determines the algorithm that is used during training.

Arguments

There are two types of ratings (user feedback): EXPLICIT and IMPLICIT. Use the one that best fits your use case.

  • If the user has explicitly provided a rating, for example 1-5, to an item such as movie recommendations, then specify EXPLICIT. This trains the model using the Alternating Least Squares algorithm. This is the default value.

  • If you don't have explicit user feedback, the rating value must be artificially constructed based on the user's interaction with the item, for example, by looking at their clicks, pageviews, and purchases. In this situation, specify IMPLICIT. This trains the model using the Weighted-Alternating Least Squares algorithm.

For more information about the differences between the two feedback types and when to use which type, see Feedback types.

NUM_FACTORS

Syntax

NUM_FACTORS = { int64_value | HPARAM_RANGE(range) | HPARAM_CANDIDATES([candidates]) }

Description

Specifies the number of latent factors to use.

Arguments

If you aren't running hyperparameter tuning, then you can specify an INT64 value between 2 and 200. The default value is log2(n), where n is the number of training examples.

If you are running hyperparameter tuning, use one of the following options:

  • The HPARAM_RANGE keyword and two INT64 values that define the range of the hyperparameter. For example, NUM_FACTORS = HPARAM_RANGE(5, 20).
  • The HPARAM_CANDIDATES keyword and an array of INT64 values that provide discrete values to use for the hyperparameter. For example, NUM_FACTORS = HPARAM_CANDIDATES([5, 10, 20, 40]).

When running hyperparameter tuning, the valid range is [2, 200],the default range is [2, 20], and the scale type is LINEAR.

USER_COL

Syntax

USER_COL = string_value

Description

The user column name.

Arguments

A STRING value. The default value is user.

ITEM_COL

Syntax

ITEM_COL = string_value

Description

The item column name.

Arguments

A STRING value. The default value is item.

RATING_COL

Syntax

RATING_COL = string_value

Description

The rating column name.

Arguments

A STRING value. The default value is rating.

WALS_ALPHA

Syntax

WALS_ALPHA = { float64_value | HPARAM_RANGE(range) | HPARAM_CANDIDATES([candidates]) }

Description

Adjusts the user feedback's impact on recommendation confidence, balancing positive and negative input in the loss function.

You can only use this hyperparameter with IMPLICIT matrix factorization models.

For more information, see Feedback types.

Arguments

If you aren't running hyperparameter tuning, then you can specify a FLOAT64 value. The default value is 40.0.

If you are running hyperparameter tuning, use one of the following options:

  • The HPARAM_RANGE keyword and two FLOAT64 values that define the range of the hyperparameter. For example, WALS_ALPHA = HPARAM_RANGE(0, 5.0).
  • The HPARAM_CANDIDATES keyword and an array of FLOAT64 values that provide discrete values to use for the hyperparameter. For example, WALS_ALPHA = HPARAM_CANDIDATES([0, 1.0, 3.0, 5.0]).

When running hyperparameter tuning, the valid range is (0, ∞], the default range is [0, 100.0], and the scale type is LINEAR.

L2_REG

Syntax

L2_REG = { float64_value | HPARAM_RANGE(range) | HPARAM_CANDIDATES([candidates]) }

Description

The amount of L2 regularization applied.

Arguments

If you aren't running hyperparameter tuning, then you can specify a FLOAT64 value. The default value is 1.0.

If you are running hyperparameter tuning, then you can use one of the following options:

  • The HPARAM_RANGE keyword and two FLOAT64 values that define the range to use for the hyperparameter. For example, L2_REG = HPARAM_RANGE(1.5, 5.0).
  • The HPARAM_CANDIDATES keyword and an array of FLOAT64 values that provide discrete values to use for the hyperparameter. For example, L2_REG = HPARAM_CANDIDATES([0, 1.0, 3.0, 5.0]).

When running hyperparameter tuning, the valid range is (0, ∞), the default range is (0, 10.0], and the scale type is LOG.

MAX_ITERATIONS

Syntax

MAX_ITERATIONS = int64_value

Description

The maximum number of training iterations, where one iteration represents a single pass of the entire training data.

Arguments

An INT64 value. The default value is 20.

EARLY_STOP

Syntax

EARLY_STOP = { TRUE | FALSE }

Description

Determines whether training should stop after the first iteration in which the relative loss improvement is less than the value specified for MIN_REL_PROGRESS.

Arguments

A BOOL value. The default value is TRUE.

MIN_REL_PROGRESS

Syntax

MIN_REL_PROGRESS = float64_value

Description

The minimum relative loss improvement that is necessary to continue training when EARLY_STOP is set to TRUE. For example, a value of 0.01 specifies that each iteration must reduce the loss by 1% for training to continue.

Arguments

A FLOAT64 value. The default value is 0.01.

DATA_SPLIT_METHOD

Syntax

DATA_SPLIT_METHOD = { 'AUTO_SPLIT' | 'RANDOM' | 'CUSTOM' | 'SEQ' | 'NO_SPLIT' }

Description

The method used to split input data into training, evaluation, and, if you are running hyperparameter tuning, test data sets. Training data is used to train the model. Evaluation data is used to avoid overfitting by using early stopping. Test data is used to test the hyperparameter tuning trial and record its metrics in the model.

The percentage sizes of the data sets produced by the various arguments for this option are approximate. Larger input data sets come closer to the percentages described than smaller input data sets do.

You can see the model's data split information in the following ways:

  • The data split method and percentage are shown in the Training Options section of the model's Details page on the BigQuery page of the Google Cloud console.
  • Links to temporary tables that contain the split data are available in the Model Details section of the model's Details page on the BigQuery of the Google Cloud console. You can also return this information from the DataSplitResult field in the BigQuery API. These tables are saved for 48 hours. If you need this information for more than 48 hours, then you should export this data or copy it to permanent tables.

Arguments

This option accepts the following values:

* AUTO_SPLIT: This option splits the data as follows:
  • If there are fewer than 500 rows in the input data, then all rows are used as training data.
  • If you aren't running hyperparameter tuning, then data is randomized and split as follows:

    • If there are between 500 and 50,000 rows in the input data, then 20% of the data is used as evaluation data and 80% is used as training data.
    • If there are more than 50,000 rows, then 10,000 rows are used as evaluation data and the remaining rows are used as training data.
  • If you are running hyperparameter tuning and there are more than 500 rows in the input data, then the data is randomized and split as follows:

    • 10% of the data is used as evaluation data
    • 10% is used as test data
    • 80% is used as training data

      For more information, see Data split.

  • RANDOM: Data is randomized before being split into sets. You can use this option with the DATA_SPLIT_EVAL_FRACTION and DATA_SPLIT_TEST_FRACTION options to customize the data split. If you don't specify either of those options, data is split in the same way as for the AUTO_SPLIT option.

    A random split is deterministic: different training runs produce the same split results if the same underlying training data is used.

  • CUSTOM: Split data using the value in a specified column:

    • If you aren't running hyperparameter tuning, then you must provide the name of a column of type BOOL. Rows with a value of TRUE or NULL are used as evaluation data, rows with a value of FALSE are used as training data.
    • If you are running hyperparameter tuning, then you must provide the name of a column of type STRING. Rows with a value of TRAIN are used as training data, rows with a value of EVAL are used as evaluation data, and rows with a value of TEST are used as test data.

    Use the DATA_SPLIT_COL option to identify the column that contains the data split information.

  • SEQ: Split data sequentially by using the value in a specified column of one of the following types:

    • NUMERIC
    • BIGNUMERIC
    • STRING
    • TIMESTAMP

    The data is sorted smallest to largest based on the specified column.

    When you aren't running hyperparameter tuning, the first n rows are used as evaluation data, where n is the value specified for DATA_SPLIT_EVAL_FRACTION. The remaining rows are used as training data.

    When you are running hyperparameter tuning, the first n rows are used as evaluation data, where n is the value specified for DATA_SPLIT_EVAL_FRACTION. The next m rows are used as test data, where m is the value specified for DATA_SPLIT_TEST_FRACTION. The remaining rows are used as training data.

    All rows with split values smaller than the threshold are used as training data. The remaining rows, including NULLs, are used as evaluation data.

    Use the DATA_SPLIT_COL option to identify the column that contains the data split information.

  • NO_SPLIT: This is the default. No data split; all input data is used as training data.

    While the other methods are supported, use them with caution. Due to the nature of the matrix factorization algorithm, if a split eliminates all of the ratings for a user or item, a factor weight vector is not generated for the user or item.

    However, use this option with caution when tuning the NUM_FACTORS hyperparameter. Although NO_SPLIT allows better performance when using larger values for NUM_FACTORS, the validity of the results might be degraded.

DATA_SPLIT_EVAL_FRACTION

Syntax

DATA_SPLIT_EVAL_FRACTION = float64_value

Description

The fraction of the data to use as evaluation data. Use when you are specifying RANDOM or SEQ as the value for the DATA_SPLIT_METHOD option.

If you are running hyperparameter tuning and you specify a value for this option, you must also specify a value for DATA_SPLIT_TEST_FRACTION. In this case, the training dataset is 1 - eval_fraction - test_fraction. For example, if you specify 20.00 for DATA_SPLIT_EVAL_FRACTION and 8.0 for DATA_SPLIT_TEST_FRACTION, your training dataset is 72% of the input data.

Arguments

A FLOAT64 value. The default is 0.2. The service maintains the accuracy of the input value to two decimal places.

DATA_SPLIT_TEST_FRACTION

Syntax

DATA_SPLIT_TEST_FRACTION = float64_value

Description

The fraction of the data to use as test data. Use this option when you are running hyperparameter tuning and specifying either RANDOM or SEQ as value for the DATA_SPLIT_METHOD option.

If you specify a value for this option, you must also specify a value for DATA_SPLIT_EVAL_FRACTION. In this case, the training dataset is 1 - eval_fraction - test_fraction. For example, if you specify 20.00 for DATA_SPLIT_EVAL_FRACTION and 8.0 for DATA_SPLIT_TEST_FRACTION, your training dataset is 72% of the input data.

Arguments

A FLOAT64 value. The default is 0. The service maintains the accuracy of the input value to two decimal places.

DATA_SPLIT_COL

Syntax

DATA_SPLIT_COL = string_value

Description

The name of the column to use to sort input data into the training, evaluation, or test set. Use when you are specifying CUSTOM or SEQ as the value for the DATA_SPLIT_METHOD option:

  • If you aren't running hyperparameter tuning and you are specifying SEQ as the value for DATA_SPLIT_METHOD, then the data is first sorted smallest to largest based on the specified column. The last n rows are used as evaluation data, where n is the value specified for DATA_SPLIT_EVAL_FRACTION. The remaining rows are used as training data.
  • If you aren't running hyperparameter tuning and you are specifying CUSTOM as the value for DATA_SPLIT_METHOD, then you must provide the name of a column of type BOOL. Rows with a value of TRUE or NULLare used as evaluation data, rows with a value of FALSE are used as training data.
  • If you are running hyperparameter tuning and you are specifying SEQ as the value for DATA_SPLIT_METHOD, then the data is first sorted smallest to largest based on the specified column. The last n rows are used as evaluation data, where n is the value specified for DATA_SPLIT_EVAL_FRACTION. The next m rows are used as test data, where m is the value specified for DATA_SPLIT_TEST_FRACTION. The remaining rows are used as training data.
  • If you are running hyperparameter tuning and you are specifying CUSTOM as the value for DATA_SPLIT_METHOD, then you must provide the name of a column of type STRING. Rows with a value of TRAIN are used as training data, rows with a value of EVAL are used as evaluation data, and rows with a value of TEST are used as test data.

The column you specify for DATA_SPLIT_COL can't be used as a feature or label, and is excluded from features automatically.

Arguments

A STRING value.

NUM_TRIALS

Syntax

NUM_TRIALS = int64_value

Description

The maximum number of submodels to train. The tuning stops when NUM_TRIALS submodels are trained, or when the hyperparameter search space is exhausted. You must specify this option in order to use hyperparameter tuning.

Arguments

An INT64 value between 1 and 100, inclusive.

MAX_PARALLEL_TRIALS

Syntax

MAX_PARALLEL_TRIALS = int64_value

Description

The maximum number of trials to run at the same time. If you specify a value for this option, you must also specify a value for NUM_TRIALS.

Arguments

An INT64 value between 1 and 5, inclusive. The default value is 1.

HPARAM_TUNING_ALGORITHM

Syntax

HPARAM_TUNING_ALGORITHM = { 'VIZIER_DEFAULT' | 'RANDOM_SEARCH' | 'GRID_SEARCH' }

Description

The algorithm used to tune the hyperparameters. If you specify a value for this option, you must also specify a value for NUM_TRIALS.

Arguments

Specify one of the following values:

  • VIZIER_DEFAULT: Use the default algorithm in Vertex AI Vizier to tune hyperparameters. This algorithm is the most powerful algorithm of those offered. It performs a mixture of advanced search algorithms, including Bayesian optimization with Gaussian processes. It also uses transfer learning to take advantage of previously tuned models. This is the default, and also the recommended approach.

  • RANDOM_SEARCH: Use random search to explore the search space.

  • GRID_SEARCH: Use grid search to explore the search space. You can only use this algorithm when every hyperparameter's search space is discrete.

HPARAM_TUNING_OBJECTIVES

Syntax

For explicit MATRIX_FACTORIZATION models:

HPARAM_TUNING_OBJECTIVES = 'MEAN_SQUARED_ERROR'

For implicit MATRIX_FACTORIZATION models:

HPARAM_TUNING_OBJECTIVES = { 'MEAN_AVERAGE_PRECISION' | 'MEAN_SQUARED_ERROR' | 'NORMALIZED_DISCOUNTED_CUMULATIVE_GAIN' | 'AVERAGE_RANK' }

Description

The hyperparameter tuning objective for the model; only one objective is supported. If you specify a value for this option, you must also specify a value for NUM_TRIALS.

Arguments

The possible objectives are a subset of the model evaluation metrics for the model type. If you aren't running hyperparameter tuning, or if you are and you don't specify an objective, the default objective is used. For explicit MATRIX_FACTORIZATION models, the default is MEAN_SQUARED_ERROR'. For implicit MATRIX_FACTORIZATION models , the default is MEAN_AVERAGE_PRECISION.

MODEL_REGISTRY

The MODEL_REGISTRY option specifies the model registry destination. VERTEX_AI is the only supported model registry destination. To learn more, see Register a BigQuery ML model.

VERTEX_AI_MODEL_ID

The VERTEX_AI_MODEL_ID option specifies the Vertex AI model ID to register the model with.

You can only set the VERTEX_AI_MODEL_ID option when the MODEL_REGISTRY option is set to VERTEX_AI. To learn more, see Add a Vertex AI model ID.

VERTEX_AI_MODEL_VERSION_ALIASES

The VERTEX_AI_MODEL_VERSION_ALIASES option specifies the Vertex AI model alias to register the model with.

You can only set the VERTEX_AI_MODEL_VERSION_ALIASES option when the MODEL_REGISTRY option is set to VERTEX_AI. To learn more, see Add a Vertex AI model ID.

KMS_KEY_NAME

Syntax

KMS_KEY_NAME = string_value

Description

The Cloud Key Management Service customer-managed encryption key (CMEK) to use to encrypt the model.

Arguments

A STRING value containing the fully-qualified name of the CMEK. For example,

'projects/my_project/locations/my_location/keyRings/my_ring/cryptoKeys/my_key'

query_statement

The AS query_statement clause specifies the GoogleSQL query used to generate the training data. See the GoogleSQL query syntax page for the supported SQL syntax of the query_statement clause.

The query_statement is expected to contain exactly 3 columns (user, item, and rating) unless you specify a DATA_SPLIT_METHOD value that requires use of a DATA_SPLIT_COL value.

Supported inputs

BigQuery supports different GoogleSQL data types for the input columns for matrix factorization. Supported data types for each respective column include:

Matrix factorization input column Supported types
user Any groupable data type
item Any groupable data type
rating INT64
NUMERIC
BIGNUMERIC
FLOAT64

Feedback types

An important part of creating a good matrix factorization model for recommendations is to make sure that data is trained on the algorithm that is best suited for it. For matrix factorization models, there are two different ways to get a rating for a user-item pair.

Ratings provided by the user are considered to be explicit feedback. A low explicit rating tends to imply the user felt very negatively about an item while a high explicit rating tends to imply that the use liked the item. Movie streaming sites where users give ratings are examples of explicitly labeled datasets. For explicit feedback problems, BigQuery ML uses the alternating least squares algorithm (ALS). ALS seeks to minimize the following loss function:

$$ Loss = \sum_{u, i \in \text{observed ratings}} (r_{ui} - x^T_uy_i)^2 + \lambda(\sum_u||x_u||^2 + \sum_i||y_i||^2)$$

Where

\(r_{ui} = \) rating that user \(u\) gave to item \(i\)
\(x_u = \) latent factor weights vector for user \(u\). Is length NUM_FACTORS.
\(y_i = \) latent factor weights vector for item \(i\). Is length NUM_FACTORS.
\(\lambda = \) L2_REG

However, most of the time data isn't labeled by users. Often, the only metrics that you have as to whether a user liked an item or movie is by the click rate or engagement time. You can use this as a proxy rating, but it is not necessarily a definitive indication as to whether a user likes or dislikes something. The data in these datasets is considered to be implicit feedback. For implicit feedback problems, BigQuery ML uses a variant of the ALS algorithm called weighted-alternating least squares (WALS), which is described in http://yifanhu.net/PUB/cf.pdf. This approach uses these proxy ratings and treats them as an indicator of the interest that a user has in an item. WALS seeks to minimize the following loss function:

$$ Loss = \sum_{u, i} c_{ui}(p_{ui} - x^T_uy_i)^2 + \lambda(\sum_u||x_u||^2 + \sum_i||y_i||^2) $$

Where, in addition to the variables defined above, the function also introduces the following variables:

\(p_{ui} = 1\) when \(r_{ui} > 0\) and \(p_{ui} = 0\) when \(r_{ui} < 0\)
\(c_{ui} = 1 + \alpha r_{ui}\)
\(\alpha = \) WALS_ALPHA

For explicit matrix factorization, the input is typically integers within a known fixed range. For implicit matrix factorization, the input ratings can be doubles or integers that span a wider range. We recommend that you make sure there aren't any outliers in the input ratings, and that you scale the input ratings if the model is performing poorly.

Known limitations

If you get the "Model is too large (>100 MB)" error, check the input data. This error is caused by having too many ratings for a single user or a single item. Hashing the user or item columns into an INT64 value or reducing the data size can help. You can use the following formula to determine whether this error might occur:

max(num_rated_user, num_rated_item) < 100 million

Where num_rated_user is the maximum item ratings that a single user has entered and num_rated_items is the maximum user ratings for a given item.

Examples

The following example creates models named mymodel in dataset mydataset in your default project.

Train a matrix factorization model with explicit feedback

This example creates an explicit feedback matrix factorization model.

CREATE MODEL `project_id.mydataset.mymodel`
 OPTIONS(MODEL_TYPE='MATRIX_FACTORIZATION') AS
SELECT
  user,
  item,
  rating
FROM
  `mydataset.mytable`

Train a matrix factorization model with implicit feedback

This example creates an implicit feedback matrix factorization model.

CREATE MODEL `project_id.mydataset.mymodel`
 OPTIONS(MODEL_TYPE='MATRIX_FACTORIZATION',
         FEEDBACK_TYPE='IMPLICIT') AS
SELECT
  user,
  item,
  rating
FROM
  `mydataset.mytable`

What's next