What's changed: Created Professional Machine Learning Engineer Chapter 1 (Domain 1 "Low-code": BigQuery ML (classification/regression/ARIMA_PLUS/matrix factorization/boosted trees/autoencoders, features, ML.PREDICT); AutoML (data prep/labeling/Tabular Workflows/tabular-text-image-video-forecasting/debug); Model Garden ML APIs; industry-specific APIs (Document AI/Retail); RAG and Vertex AI Agent Builder).
1.1BigQuery ML and AutoML
Understand choosing the right BigQuery ML model by business problem (linear/binary classification, regression, time-series, matrix factorization, boosted trees, autoencoders), feature engineering and prediction with BigQuery ML, and AutoML (data prep/labeling/Tabular Workflows, custom/forecasting models on tabular data, configuring/debugging trained models). (Note: legacy AutoML Text is being deprecated in favor of Gemini-based models on Vertex AI; AutoML for image/video/tabular continues.)
A Professional Machine Learning Engineer builds, trains, productionizes, and optimizes AI on Google Cloud. The first pillar is low-code: the judgment to realize ML by the shortest route fitting the data's location and the problem—without writing much code.
1.1.1BigQuery ML
If data is in BigQuery, the shortest path is BigQuery ML (BQML)—train/predict in SQL without moving data. Choose the model by problem: classification = logistic regression, numeric prediction = linear regression, time-series (e.g., demand) = ARIMA_PLUS, recommendations = matrix factorization, high-accuracy tabular = boosted trees (XGBoost), anomaly detection = autoencoders, etc. Do feature engineering (embed preprocessing in training via the TRANSFORM clause) and prediction with ML.PREDICT—all in SQL. Map "data in BigQuery, done in SQL = BigQuery ML."
1.1.2AutoML
For a high-quality custom model without code, use AutoML (Vertex AI). From tabular, image, or video data (text/speech have moved to Gemini / the Speech API), Google automatically searches architectures and tunes hyperparameters. Prepare data with labeling and feature selection, and for advanced tabular control use Tabular Workflows. For demand/inventory, build an AutoML forecasting model, then configure/debug using metrics after training. Choose "no-code high-quality custom model = AutoML" vs "quick in SQL = BigQuery ML."
Common: problem/data → means. E.g., "data in BigQuery, quick classification/regression in SQL" = BigQuery ML; "time-series demand forecast in SQL" = BigQuery ML ARIMA_PLUS; "recommendations" = matrix factorization; "no-code custom model for image/text" = AutoML; "advanced tabular control" = Tabular Workflows; "need labeled data" = data labeling.
Watch the mix-ups: (1) Choose BigQuery ML (data in BigQuery, SQL) vs AutoML (diverse data types, no-code GUI/SDK) by problem. (2) Time-series = ARIMA_PLUS; recommendations = matrix factorization—match the model type to the problem. (3) AutoML costs time/money to train—for small/instant needs consider BQML or pretrained APIs.
1.1.3Section summary
- Data in BigQuery, SQL-only = BigQuery ML (classification/regression/ARIMA_PLUS/matrix factorization/boosted trees)
- No-code high-quality custom models on diverse data = AutoML (Tabular Workflows/forecasting)
- Choose the model type by problem (classification/regression/time-series/recommendation/anomaly)
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Data is already in BigQuery and you want a quick binary classification model and predictions using only SQL. Which is best?
Q2. To build a high-quality custom image classification model without writing code, which is best?
Q3. To build a time-series demand forecasting model in SQL on BigQuery, which model?
Q4. To build recommendations from user-item interactions, which BigQuery ML model?
Q5. For more advanced control (feature transforms, architecture tuning) with AutoML on tabular data, which is best?