Semantic-Similarity recipe
The Semantic-Similarity recipe (aws-semantic-similarity) generates recommendations for items that are semantically similar to a given item based on textual content. Unlike traditional similarity recipes that rely on user-item interactions, this recipe analyzes the textual descriptions and attributes of items to generate embeddings and identify semantically similar items
This recipe is ideal for scenarios where you want to recommend items based on content similarity, such as recommending books with similar themes, articles on related topics, or products with similar descriptions. It works particularly well for new items with limited interaction history (cold-start scenarios) and for catalogs where semantic relationships are more important than co-occurrence patterns.
With Semantic-Similarity, you provide an item ID in a GetRecommendations operation (or the Amazon Personalize console) and Amazon Personalize returns a list of similar items. Or you can use a batch workflow to get similar items for all of the items in your inventory (see Getting batch item recommendations).
Recipe features
Semantic-Similarity uses the following Amazon Personalize recipe features when generating item recommendations:
-
Real-time personalization – With the Semantic-Similarity recipe, Amazon Personalize automatically keeps your item catalog up to date. When you add new items to your Items dataset or update existing item metadata, these changes are reflected in your recommendations within approximately 30 minutes when using incremental training. This ensures that your customers always see the most current items available in your catalog without requiring manual intervention or waiting for a full retraining cycle. This is particularly valuable for catalogs that change frequently, such as news articles, blog posts, or seasonal product offerings. To enable incremental updates, customers must:
-
Set
performIncrementalUpdatetotruefor the solution in the API -
Choose either "Full and Incremental training" or "Incremental training" option under Training method in the UI
Note that enabling incremental updates will incur additional costs whenever an update is being performed.
-
-
Metadata with recommendations – With the Semantic-Similarity recipe, campaigns automatically have the option to include item metadata with recommendation results. You don't have manually enable metadata for your campaign. You might use metadata to enrich recommendations in your user interface, such as adding the genres for movies to carousels. For more information, see Item metadata in recommendations.
Required and optional datasets
To use the Semantic-Similarity recipe, you must create an Items dataset. Amazon Personalize generates recommendations based on the semantic meaning of item metadata. For more information, see Item metadata. Semantic-Similarity can train on up to 10 million items in your Items dataset.
With Semantic-Similarity, Amazon Personalize requires Items data that includes the following
-
itemName field – One string field with
itemNameset totrue. This field should contain the title or name of the item. -
Textual description field – At least one string field marked as
textualthat contains the description of the item. This should be the field that best describes and represents the item.
Amazon Personalize uses this field to generate semantic embeddings that capture the meaning and content of your items.
Additionally, the reserved CREATION_TIMESTAMP field should be set if you want to use freshness-based ranking. For more information, see Properties and hyperparameters.
The following datasets are optional and can improve recommendations:
-
Interactions dataset– Amazon Personalize can use data in your Interactions dataset to calculate popularity scores based on user engagement with items. You can use popularity scores to rank similar items by how popular they are among users. You must provide an Interactions dataset if you want to use popularity-based ranking. You can also use data in an Interactions dataset to filter recommendations. For information about the interaction data you can import, see Item interaction data
Properties and hyperparameters
The Semantic-Similarity recipe has the following properties:
-
Name –
aws-semantic-similarity -
Recipe Amazon Resource Name (ARN) –
arn:aws:personalize:::recipe/aws-semantic-similarity -
Algorithm ARN –
arn:aws:personalize:::algorithm/aws-semantic-similarity -
Feature transformation ARN –
arn:aws:personalize:::feature-transformation/aws-semantic-similarity -
Recipe type –
RELATED_ITEMS
For more information, see Choosing a recipe.
The Semantic-Similarity recipe has no exposed hyperparameters but you can configure popularity and freshness factors when you create a campaign to influence the ranking of similar items.
The table provides the following information for each factor:
-
Range: [lower bound, upper bound]
-
Value type: Integer, Continuous (float), Categorical (Boolean, list, string)
| Name | Description |
|---|---|
| Freshness |
The freshness factor represents how recent an item is. Freshness is computed by normalizing the age of the item based on its CREATION_TIMESTAMP. To use the freshness factor, you must include the CREATION_TIMESTAMP field in your Items dataset schema. Higher values of freshness factor will prioritize newer items among semantically similar recommendations Default value: Range: Value type: Double |
| Popularity |
The popularity factor represents how popular an item is based on user interactions. Popularity is computed by normalizing the number of interactions each item received. To use the popularity factor, you must include an Interactions dataset when creating your dataset group. Higher values of popularity factor prioritize items with more customer interactions among semantically similar recommendations. Default value: Range: Value type: Double |
Note that freshness and popularity scores are computed at training and incremental updates will not update popularity and freshness scores. For the most recent popularity and freshness factors to influence the ranking of recommended items, either use automatic retraining or manually retrain the solution and update the campaign with the new solution version.
