For greater control over your ranking strategies, you can incorporate custom data you've imported into rank models or solution targeting and override rules. A model feature can be a product, user, or query metric or attribute, such as buy-to-detail rate, color, or age.
For more information on importing data for use in model features, see our guide here.
Table of Contents
- Understanding Model Features
- Creating Custom Model Features
- Using Model Features in Rank Models
- Using Model Features in Solution Targeting & Override Rules
Understanding Model Features
At its core, a model feature is a reference to a field in a data source. Model features can have two types of sources:
- Dataset
- Model Features with a dataset source point a dataset that has been imported to Convert.
- In addition to the repository and dataset, model features must define the column within the dataset that should be used as the data source.
- The ID field in the dataset references the product, user, or item being described by the model feature, while the model feature field supplies the value for that object.
- Runtime
- Model Features with a runtime source receive data values from the Relevance request that Convert receives from the customer website.
- Runtime model features can only be used for solution targeting and override rules.
- The Name of runtime model features must match the parameter name sent in the Relevance request exactly.
Convert supports five types of model features:
- Product
- A product feature is a metric or attribute related to a product, such as buy-to-detail rate, color, or date added.
- User
- A user feature is a metric or attribute related to a user, such as age, gender, or location.
- Query
- A query feature is a metric or attribute related to the query being viewed (a query is typically the category or search term being viewed), such as date created, related queries, and more.
- Query-Item
- A query-item feature is a metric or attribute that may have a unique value for each query-item pair. This could include things like list clickthrough rate, date added, and more.
- User-Item
- A user-item feature is a metric or attribute that may have a unique value for each user-item pair. This could include things like previous purchases, number of views, and more.
Creating Custom Model Features
- Navigate to the Model Features page.
- Click Add Model Feature.
- Select the desired Source.
- If the source selected is Dataset, enter a unique, identifiable Name.
- If the source selected is Runtime, enter the corresponding Name passed in the Relevance request, then jump to step 8.
- If the source selected is Dataset, use the dropdown to select the appropriate Repository.
- If the source selected is Dataset, use the dropdown to select the Dataset that contains your feature data.
- If the source selected is Dataset, enter the exact Field Name from your dataset to identify the data to be used.
To reduce potential errors, field names are not case sensitive. However, if you're uncertain what column header is used in the dataset, we recommend exporting it to confirm.
- Select the appropriate Type.
- If you plan on using the feature in rank models, enable Available for Model Training.
- This field will not be available if the Source is Dataset.
- If you plan on using the feature in solution targeting or override rules, enable Available for Targeting & Override Rules.
- Click Create Model Feature.
Once the feature has been created, all settings can be adjusted as needed.
Using Model Features in Rank Models
For a rank model to be able to use a model feature, the model feature must have Available for Model Training enabled.
Please note that creating adding a new model feature to rank models isn't always as simple as uploading a spreadsheet with the color of each product in your catalog, creating a model feature, and adding it to a rank model. To be useful to Convert's ranking models, a model feature should be in a specific numerical format. Making categorical attributes (such as color) useful to rank models requires a complex transformation process that may require a background in data science or machine learning to be effective. For assistance, or more information, please contact Convert Support.
However, metric-based custom data can be added to rank models without additional transformation.
For instructions on adding custom model features to a ranking model, view our guide on Managing Ranking Models.
Using Model Features in Solution Targeting & Override Rules
For a Solution to be able to use a model feature in targeting or override rules, the model feature must have Available for Targeting & Override Rules enabled.
Targeting rules can only use model features where the type is user or query.
Override rule result criteria can only use model features where the type is item, user-item, or query-item. Override rule context criteria can only use model features where the type is user or query.
If a model feature used in a targeting or override rule has a source of Runtime, the value provided by the relevance request will be analyzed to see if the rule applies.
If the source of the model feature is Dataset, Convert will look up the appropriate row by the ID associated with the object, then analyze the value to see if the rule applies. For example, if a rule calls for override products where the color equals red, Convert will take the product ID from the relevance request, find the matching row in the dataset, and evaluate the rule based on the value provided in the column defined by the model feature.
For instructions on managing solution targeting & override rules, view our guide on Managing Solutions & Campaigns.