Since we last blipped about Google BigQuery ML, more sophisticated models such as Deep Neural Networks and AutoML Tables have been added by connecting BigQuery ML with TensorFlow and Vertex AI as its backend. BigQuery has also introduced support for time series forecasting. One of our concerns previously was explainability. Earlier this year, BigQuery Explainable AI was announced for general availability, taking a step in addressing this. We can also export BigQuery ML models to Cloud Storage as a Tensorflow SavedModel and use them for online prediction. There remain trade-offs like ease of "continuous delivery for machine learning" but with its low barrier to entry, BigQuery ML remains an attractive option, particularly when the data already resides in BigQuery.
Often training and predicting outcomes from machine learning models require code to take the data to the model. Google BigQuery ML inverts this by bringing the model to the data. Google BigQuery is a data warehouse designed to serve large-scale queries using SQL, for analytical use cases. Google BigQuery ML extends this function and its SQL interface to create, train and evaluate machine learning models using its data sets; and eventually run model predictions to create new BigQuery data sets. It supports a limited set of models out of the box, such as linear regression for forecasting or binary and multiclass regression for classification. It also supports, with limited functionality, importing previously trained TensorFlow models. Although BigQuery ML and its SQL-based approach lower the bar for using machine learning to make predictions and recommendations, particularly for quick explorations, this comes with a difficult trade-off: compromising on other aspects of model training such as ethical bias testing, explainability and continuous delivery for machine learning.