How to train a binary classifier
A binary classifier sorts items into two groups, based off item
To get started, prepare a single table with one item per row, and
columns as attributes.
Then find the
button on the Models page, and follow along below to train an
Train a New Binary Classifier
There's a button like this on the Models page. That's the one you want.
Give your creation a name
Don't worry, you can change it later.
The name of the connection you set up earlier
This is where the training job will go for data.
The table with training data
schema.table here to specify a schema.
Make sure the user specified by the connection has permission to run
SELECT on the table.
SQL expression used to filter down items
All rows that evaluate to
true will be used to build
This is a great place to specify (for example) a date range, to get
a repeatable set of items even as the table grows.
SQL expression specifying if an item is in the test set
This will be used to split items into two groups, one for training
the model, and one for evaluating (or "testing") the model. All
rows that evaluate to
true will be used in the test
A larger training set will generally yield better models, while a
larger test set will give a more-accurate measure of model
accuracy, to a point. When in doubt, start with a 50/50 split.
Column name (or SQL expression) for the item attribute the model
This should be valued
Column name (or SQL expression) uniquely identifying the item
You'll use this later to inspect predictions for specific items.
Column names (or SQL expressions) of item attributes to use when
making a prediction
One predictor per line. There are two kinds here:
1. Numeric predictors
Real-valued attributes with many possible values, and real-valued
attributes where nearby values are related
The model will make use of the distance between attribute values.
2. Categorical predictors
All other attributes
The model will learn separately for each attribute value.
Start the training job
Your model will be ready in a couple of minutes.