A/B testing with CloudWatch Evidently - Amazon Personalize

A/B testing with CloudWatch Evidently

After you create a recommender or deploy a custom solution version with a campaign, you can perform A/B tests with Amazon Personalize recommendations and Amazon CloudWatch Evidently. The following video describes the process of using CloudWatch Evidently to perform A/B testing with Amazon Personalize recommendations. For step-by-step instructions, see Performing an A/B test with CloudWatch Evidently.

Performing an A/B test with CloudWatch Evidently

To perform an A/B test with Amazon Personalize and Amazon CloudWatch Evidently, create a CloudWatch Evidently project, define a feature and its variations, update your application to support your experiment, and create and run the experiment. As the experiment runs, you can view results in CloudWatch Evidently.

To perform an A/B test with Amazon Personalize and CloudWatch Evidently
  1. Create a CloudWatch Evidently project. A project is a logical grouping of CloudWatch resources. Within the project, you create features that have variations that you want to test or launch. For step-by-step instructions, see Create a new project in the Amazon CloudWatch User Guide.

  2. Add a feature to your project and define its variations. For this experiment, your feature should represent the recommendation scenario that you want to test, such as the click-through rate.

    When you add a feature, specify identifiers to map the different variations of your scenario to Amazon Personalize recommenders or custom campaigns. For each variation, specify the Variation type, such as String, give the variation a name, and give it a value.

    When your experiment runs, your application uses the value of variation to determine what Amazon Personalize resource to use for recommendations. For example, if you're testing two VIDEO_ON_DEMAND recommenders, one created for the Top picks for you use case and one created for the Trending now use case, you might set the following JSON as the Value for each variation.

    {"type":"top-picks-recommendations","arn":"arn:aws:personalize:us-west-2:<acct-id>:recommender/top-picks-recommender"}
    {"type":"trending-recommendations","arn":"arn:aws:personalize:us-west-2:<acct-id>:recommender/trending-now-recommender"}

    You can specify any identifier, as long as your application can use it to identify the relevant resource. For example, you might specify only the name of the recommender or campaign, and construct the Amazon Resource Name (ARN) of the resource in your application.

    For step-by-step instructions to add a feature, see Add a feature to a project in the Amazon CloudWatch User Guide.

  3. Update your application to support your experiment:

    • Feature evaluation – Use the CloudWatch Evidently EvaluateFeature API operation to assign variations to each user session. The EvaluateFeature response includes the variation value that you specified in the previous step. In this case, it's a JSON object with the type of recommender and it's the ARN of the recommender. Update your recommendation request code to get recommendations from this resource.

      For information about evaluating a feature, see Using EvaluateFeature in the Amazon CloudWatch User Guide.

    • Record outcomes – Add code to your application to track results from users' interactions with recommendations.

      To track metrics for your experiments in CloudWatch Evidently, use the CloudWatch Evidently PutProjectEvents API operation to record outcomes for each user. For example, if a user in an experiment clicks a recommended item, you would send details for this event to CloudWatch Evidently.

      For information about sending events to CloudWatch Evidently, see Using PutProjectEvents in the Amazon CloudWatch User Guide.

      To improve Amazon Personalize recommendation relevance, you can record outcome events with the Amazon Personalize PutEvents API operation. If your domain use case or custom recipe supports real-time updates to recommendations, Amazon Personalize can learn from your user’s most recent activity and update recommendations as they use your application. If it doesn't support updates, Amazon Personalize uses this data during the next full retraining of your model and then it impacts recommendations.

      For information about streaming events to Amazon Personalize, see Recording real-time events to influence recommendations.

  4. Create and start an experiment. When you create an experiment, specify the following:

    • Feature – Choose the feature to be tested in the experiment.

    • Audience – Configure how many of your users will participate, and configure how to split traffic between feature variations.

    • Metrics – Specify the metrics that determine the success of the experiment. For example, the number of clicks.

    After you finish creating the experiment, specify its duration and start the experiment. For step-by-step instructions to create and start experiments in CloudWatch Evidently, see Create an experiment in the Amazon CloudWatch User Guide.

  5. As you run your experiment, you can view results in the CloudWatch Evidently experiment dashboard. For information about viewing experiment results, see View experiment results in the dashboard in the Amazon CloudWatch User Guide.

Sample implementations

The following sample implementations show how to implement A/B testing with CloudWatch Evidently.