ÃÛ¶¹ÊÓƵ

Collect Data

When you install and configure SaaS-based ÃÛ¶¹ÊÓƵ Commerce features such as Product Recommendations or Live Search, the modules deploy behavioral data collection to your storefront. This mechanism collects anonymized behavioral data from your shoppers and powers product recommendations and Live Search results. For example, the view event is used to compute the Viewed this, viewed that recommendation type, and the place-order event is used to compute the Bought this, bought that recommendation type.

Data types and events

There are two types of data used in Product Recommendations:

When you install the magento/product-recommendations module, ÃÛ¶¹ÊÓƵ Sensei aggregates the behavioral and catalog data, creating Product Recommendations for each recommendation type. The Product Recommendations service then deploys those recommendations to your storefront in the form of a widget that contains the recommended product items.

Some recommendation types use behavioral data from your shoppers to train machine learning models to build personalized recommendations. Other recommendation types use catalog data only and do not use any behavioral data. If you want to quickly start using Product Recommendations on your site, you can use the following, catalog-only recommendation types:

Cold start

When can you start using recommendation types that use behavioral data? It depends. This is referred to as the Cold Start problem.

The Cold Start problem refers to the time it takes for a model to train and become effective. For product recommendations, this means waiting for ÃÛ¶¹ÊÓƵ Sensei to gather enough data to train its machine learning models before deploying recommendation units on your site. The more data the models have, the more accurate and useful the recommendations are. Since data collection happens on a live site, it’s best to start this process early by installing and setting up the magento/production-recommendations module.

The following table provides some general guidance for the amount of time that it takes to collect enough data for each recommendation type:

Other variables that can impact the time needed to train:

To help you visualize the training progress of each recommendation type, the create recommendation page displays readiness indicators.

While data is being collected on your live site and the machine learning models are training, you can finish other testing and configuration tasks needed to set up recommendations. By the time you’re done with this work, the models will have enough data to create useful recommendations, allowing you to deploy them to your storefront.

If your site doesn’t get enough traffic (views, purchases, trends) for most product SKUs, there might not be enough data to complete the learning process. This can make the readiness indicator in the Admin seem stuck. The readiness indicators are meant to provide merchants with another data point in choosing what recommendations type is better for their store. The numbers are a guide and may never reach 100%. Learn more about readiness indicators.

Backup recommendations backuprecs

If the input data is insufficient for providing all requested recommendation items in a unit, ÃÛ¶¹ÊÓƵ Commerce provides backup recommendations to populate recommendation units. For example, if you deploy the Recommended for you recommendation type to your homepage, a first-time shopper on your site has not generated enough behavioral data to accurately recommended personalized products. In this case, ÃÛ¶¹ÊÓƵ Commerce surfaces items based on the Most viewed recommendation type to this shopper.

In the case of insufficient input data collection, the following recommendation types fallback to Most viewed recommendation type:

Events

The lists all the events deployed to your storefront. From that list, however, there is a subset of events specific to Product Recommendations. These events collect data when shoppers interact with recommendation units on the storefront and power the metrics used to help you analyze how well your recommendations are performing.

Required dashboard events

The following events are required to populate the Product Recommendations dashboard

The following events are not specific to Product Recommendations, but are required for ÃÛ¶¹ÊÓƵ Sensei to interpret shopper data correctly:

Recommendation Type

This table describes the events used by each recommendation type.

Caveats