Machine Learning Models and Training

Each combination of recommendation model type and objective has unique user event and data requirements. The table below shows which recommendation model types and objectives we support based on the user events that are currently a part of the BigCommerce integration.

In the future, we’ll be enhancing this integration to include full support for all user events, which in turn unlocks all of the different model types and objectives.

✔ indicates that a feature is supported with the current integration.
* indicates that a feature is not yet supported, but will be in the future.

DISCLAIMER: Model types and placements (the page where you place the recommendation) have not all been tested. We recommend testing them before you deploy them on your live website.

Learn more about the data requirements for various models and objectives in Google’s documentation.

Model training

The following sections describe how BigCommerce loads data to your GCP account and the data used to train models for prediction requests.

Initial data load

BigCommerce sends catalog and shopper event data to your GCP account so that you can create machine learning models and serving configurations in your GCP project.
An initial load occurs once, which includes the most up-to-date catalog data, Product Detail Page View events from the past 90 days, and Purchase Complete events from the past 180 days.

After BigCommerce sends the data, you can configure the machine learning model you want to use in the Google Cloud console. You can manage models, including their serving configurations and model placements in your GCP account.

Continuous data updates

The following pages describe how BigCommerce keeps the models up to date:

• Catalog data
• Shopper events

Creating a ML model

To create your first ML model in GCP, navigate to the Models section and click Create model. You can name your model and select the model type and objective. The UI will show if you have sufficient data for various model types and objectives. 

Tuning preference frequency

We recommend using the default setting of “every three months”. This will ensure the efficacy of the model stays high, which means you don’t have to remember to manually retune the model in the future.

Filtering by attribute values

We recommend using the default value of “Auto generate tags”. This is useful for two scenarios:

• When products in your BigCommerce catalog are marked as “not visible” for the storefront, they do not appear in recommendation requests.
• When applying a filter when making product recommendation requests. For more on this topic, see the section on filtering. 

You can set this to “Do not generate tags” if you don’t need either of these capabilities. You can change this later by editing your model and changing the “Filter settings” option.