The ML.RECOMMEND function

This document describes the ML.RECOMMEND function, which lets you generate a predicted rating for every user-item row combination for a matrix factorization model. Because the input data for a matrix factorization model tends to be a sparse matrix with missing values, ML.RECOMMEND can return the predictions for those missing values without requiring specification of each entry.

Syntax

ML.RECOMMEND(
  MODEL `project_id.dataset.model`
  [, { TABLE `project_id.dataset.table` | (query_statement) }]
  [, STRUCT(trial_id AS trial_id)])

Arguments

ML.RECOMMEND takes the following arguments:

  • project_id: a STRING value that specifies your project ID.
  • dataset: a STRING value that specifies the BigQuery dataset that contains the model.
  • model: a STRING value that specifies the name of the model.

  • table: a STRING value that specifies the name of the input table that contains the user and item data.

    If you specify input data by using either the table or query_statement argument, the user and item columns must match the user and item columns in the model, and their types must be compatible according to BigQuery implicit coercion rules.

    If the input table does not contain both the user and item column, the input table can only contain one column. If the table contains both the user and item columns, then the non-user or item columns are passed through and available for use in the query.

  • query_statement: a STRING value that specifies the GoogleSQL query that is used to generate the evaluation data. For the supported SQL syntax for the query_statement clause in GoogleSQL, see Query syntax.

    If you specify input data by using either the table or query_statement argument, the user and item columns must match the user and item columns in the model, and their types must be compatible according to BigQuery implicit coercion rules.

  • trial_id: an INT64 value that identifies the hyperparameter tuning trial that you want the function to evaluate. The function uses the optimal trial by default. Only specify this argument if you ran hyperparameter tuning when creating the model.

Output

ML.RECOMMEND outputs at least 3 columns for all cases; the user column, the item column and a column for predicted recommendations.

The output of ML.RECOMMEND is computed as follows:

  • If both the user and item columns are in the input data, then ML.RECOMMEND returns a rating for each user-item pair.
  • If only the user or only the item is specified (for example, if the table identified by the table argument only contains the user column), then all the item ratings for every user in the table are outputted.
  • If either the user or item feature was not in the training dataset, the rating that is returned is the intercept of the feature that was provided, either item or user, added to the global__intercept__ offset. For example, global__intercept__ + __intercept__['user_a'].
  • If input data is specified but does not provide either the user or item column, ML.RECOMMEND returns an error.
  • If no input data is specified, ML.RECOMMEND outputs the ratings for every user and item combination seen during training.

If the model was trained with feedback_type=EXPLICIT, a user column called user, and an item column called item, then ML.RECOMMEND returns the following columns:

  • user: a STRING value containing the user data.
  • item: a STRING value containing the item data
  • predicted_<rating_col_name>: a FLOAT64 value that contains the rating for each user-item pair. Because the input ratings from training are assumed to be explicit feedback, the predicted ratings are approximately in the range of the original input, although ratings outside the range are also normal.

If the model was trained with feedback_type=IMPLICIT, a user column called user, and an item column called item, then ML.RECOMMEND returns the following columns:

  • user: a STRING value containing the user data.
  • item: a STRING value containing the item data
  • predicted_<rating_col_name>_confidence: a FLOAT64 value that contains the relative confidence for each user-item pair. The input ratings from training are assumed to be a proxy for user confidence. Therefore, if the model has converged, the predicted confidences lie between approximately 0 and 1 (but can lie just outside that range). If the model hasn't converged, the predicted confidences can be any value. If your model isn't converging and your ratings are very large, try decreasing the WALS_ALPHA value that's specified in the CREATE MODEL statement for the model. If your model isn't converging and your ratings are very small, try increasing the WALS_ALPHA value.

Examples

No input data

The following example generates predicted ratings for every user-item pair in the inputs of mymodel because there is no input data specified.

SELECT
  *
FROM
  ML.RECOMMEND(MODEL `mydataset.mymodel`)

With input data

The following example generates predicted ratings for each user-item row in mydataset.mytable assuming that mydataset.mymodel was trained using the user column user and item column item.

SELECT
  *
FROM
  ML.RECOMMEND(MODEL `mydataset.mymodel`,
      (
      SELECT
        user,
        item
      FROM
        `mydataset.mytable`))

What's next