Batch recommendations with themes from Content Generator - Amazon Personalize

Batch recommendations with themes from Content Generator


When you get batch recommendations with themes, you incur additional costs. For more information, see Amazon Personalize pricing.

If you use the Similar-Items recipe, Amazon Personalize Content Generator can add descriptive themes to batch recommendations. Content Generator is a generative artificial intelligence (generative AI) capability managed by Amazon Personalize.

When you get batch recommendations with themes, Amazon Personalize Content Generator adds a descriptive theme for each set of similar items. The theme is based on the item description and item name data in your Items dataset. Amazon Personalize includes the themes in the output of the batch inference job. You can use the themes to make the text in your application or marketing messages more compelling.

For example, if you get related items recommendations for a breakfast food item, Amazon Personalize might generate a theme like Rise and shine or Morning essentials. You might use the theme to replace a generic carousel title, like Frequently bought together. Or you might incorporate the theme in a promotional email or marketing campaign for new menu options.

AWS doesn't monitor themes from Content Generator. To confirm the theme quality, you can use the scores produced for each recommended item. For more information, see Ranking and scoring for batch recommendations with themes.

Supported regions

Amazon Personalize Content Generator is only available in the following AWS Regions:

  • US East (N. Virginia)

  • US West (Oregon)

  • Asia Pacific (Tokyo)

Guidelines and requirements

The following are guidelines and requirements for generating recommendations with themes:

Ranking and scoring for batch recommendations with themes

When you get batch recommendations with themes, Amazon Personalize ranks each set of items based on how relevant the theme is for each item. Each item includes a score in a rough range of -0.1 and 0.6. The higher the score, the more closely related the item is to the theme. You might use the scores to set a threshold to show only items that are strongly related to the theme.

For example, Amazon Personalize might return a theme of For your sweet tooth, and the related items and their scores might be: hard candy (score 0.19884521), chocolate (score .17664525), apple (score .08994528), popsicle (score .14294521), sweet potato (score .07794527), and carrot (score .04994523). In your application, you might add a rule to include only items with a score of .10 or greater, eliminating the fruits and vegetables.

The following example shows the format of the output of a batch inference job that generates movie recommendations with themes.

{"input":{"itemId":"40"},"output":{"recommendedItems":["36","50","44","22","21","29","3","1","2","39"],"theme":"Movies with a strong female lead","itemsThemeRelevanceScores":[0.19994527,0.183059963,0.17478035,0.1618133,0.1574806,0.15468733,0.1499242,0.14353688,0.13531424,0.10291852]}} {"input":{"itemId":"43"},"output":{"recommendedItems":["50","21","36","3","17","2","39","1","10","5"],"theme":"The best movies of 1995","itemsThemeRelevanceScores":[0.184988,0.1795761,0.11143453,0.0989443,0.08258403,0.07952615,0.07115086,0.0621634,-0.138913,-0.188913]}} ...

Generating batch recommendations with themes

To generate batch recommendations with themes, you complete the batch workflow as described in Batch workflow. You prepare your input data in the same way you would for a RELATED_ITEMS recipe. For an example, see RELATED_ITEMS recipes.

When you create the batch inference job, you enable theme generation and specify the item title column of your Items dataset.