BigQuery ML

Supported Model Types

Steps

1. ETL into BigQuery
2. Select and pre-process features—BigQuery does some of this, e.g. one-hot encoding
3. Create model—CREATE MODEL
   • Hyperparameters—tune model to achieve best result
   • Manual or automatic
4. Evaluate model—ML.EVALUATE
5. Get predictions—ML.PREDICT

Key Commands

• Create model: CREATE OR REPLACE MODEL
  • OPTIONS—model_type required
• Training progress: ML.TRAINING_INFO
• Inspect what a model learned: ML.WEIGHTS
  • Output: weight -1 to 1 for each feature—magnitude of weight indicates relevance of feature to model
• Stats/metrics about features/columns: ML.FEATURE_INFO
• Evaluate performance against evaluation data set: ML.EVALUATE
• Batch predictions: ML.PREDICT
  • Pass data to make predictions on

Supervised Models

• Require column called label, or specify label column in model options
• Features are columns in select statement
• Model object—object in BigQuery data set

Generative AI

• Generate text: ML.GENERATE_TEXT
  • Generates text using text-bison LLM using data from BigQuery tables