Rank models are the machine learning-based algorithms that determine the optimal sort order for a set of products.
Rank models rely on a set of model features, or a collection of product attributes and metrics, to determine which products are most likely to be purchased.
After you've created a data integration and site, the next step is to create your first rank model. Convert offers a couple basic rank model templates you can use to easily create rank models for different goals.
Once you've created a rank model, you can experiment with performance by adding custom model features. For more information on model features, view our support article here.
Table of Contents
Creating a Rank Model
- Navigate to the Rank Models page.
- Click Add Rank Model.
- Enter a unique and identifiable Name.
We recommend including the type of model you're creating for clarity. For example, if you're going to build a model that optimizes purchases, consider naming it something like "Purchases - [Descriptor]".
- Use the dropdown to select a Site.
The rank model will use generate its rankings based on behavioral data associated with the site.
- Use the dropdown to select a Scope.
Scope defines the specificity of Convert's rankings. If you want to use Convert to create category or search-term-specific rankings, select product-lists. If you use BigCommerce, or want to create a generic set of product rankings, select all-products.
- Use the dropdown to select the Parent Model.
The parent model will determine a number of settings that cannot currently be configured, such as the goal of the model. Additionally, many configurable settings can be set to inherit from the parent, and will update whenever changed in the parent.
- Click Create Rank Model.
Managing Model Features
Although you cannot currently manage the default model features derived from your Google Analytics data in a rank model, you can add and enable custom model features that you've created from third-party data sources.
Adding a Custom Model Feature
- Navigate to an existing rank model, or create a new one.
- In the Model Features tab, use the dropdown in the Name column to select a model feature. This will add it to the rank model.
If there are no empty rows available, click Add More Features at the bottom of the model features list.
- Once a feature has been added to the model, it can be enabled, disabled, or removed.
- Click Save.
Local Configuration
Once you've added a feature to your rank model, you'll need to configure whether or not the model should use it to generate its rankings. This is where the Local Configuration comes in.
Features can have one of three local configurations: Enabled, Disabled, or Inherited.
- Enabled features will be considered by the rank model when generating its rankings.
- Disabled features will not be considered by the rank model when generating its rankings.
- Inherited applies only to custom features that have been inherited from the parent model. When set to Inherited, the feature will follow whatever configuration is set on the parent model.
You can always tell what configuration is currently set on an inherited feature at the parent level by looking to the Effective Value column.
Inherited Model Features
Any custom features inherited from the parent model will appear at the top of the Model Features list. These features cannot be removed, and can be identified because their Name dropdown will be inactive.
To override the configuration of an inherited feature on the current model, use the Local Configuration field. Once you've set a local configuration on an inherited feature, any changes made to the feature's configuration at the parent level will not impact the child model. If you want an inherited feature to resume inheriting its configuration from the parent model, simply set the local configuration to Inherited.
Effective Value
For clarity, the Effective Value column will always display the feature configuration to be used by the current model.
Creating a Rank Model Version
Once you've created a rank model, it will not be ready for immediate use. First, a rank model must create a new version.
Creating a rank model version involves the model training process, where our machine learning technology processes all the data available to it, identifies and weighs key features, and ultimately establishes the algorithm to be used to rank products. Model training can take several hours.
New model versions can be created on demand or according to a schedule. To schedule regular model training, navigate to the Versioning Schedule tab within a rank model, and select your desired frequency, start time, and more. Alternatively, you click Create Version on any page within the rank model to kick off the training process on an ad hoc basis.
Default Versions
Each rank model will always have one model version set as the default version. The default version concept is designed to enable easy deployment of new rank model versions. Solutions can be configured to use the default version of a rank model, so that whenever the default version of that model changes, the solution will automatically start using the best model version.
The default version can be defined in multiple ways. To manually change the default version, navigate to the Model Versions tab, and click Make Default on the desired version. Alternatively, new model versions can automatically be marked as default at the end of the creation process. Use the Make Default Version field on the versioning schedule or manual creation form to enable this. Setting this field to If Better will only make the new version default if it has higher training scores than the current default version.
Using a Rank Model
Model Performance
Once a model version has been created, you can gain information about the model's expected performance by navigating to the Training Performance tab within the version. ERR is a sophisticated metric that quantifies the relevance of the ordering of a results set. The threshold values (@5, @10, and @20) indicate the number of products the relevance score considers. For example, ERR @20 measures the expected relevance of the first 20 products in any given list.
Note: Relevance measures such as ERR are most useful for measuring relative differences between experiences. Because of the major differences between sites and product catalogs, a "good" relevance score will vary from customer to customer. Try comparing the relevance scores of different model versions to determine which is best for your site.
Previewing Rankings
Before you can preview or implement the rankings from a rank model, you'll need to create a Solution and Campaign. Instructions for deploying rank models can be found here.
Once a rank model has been successfully deployed, you can preview its rankings via the Product Lists or All Products reports. Information on these reports can be found here.