Best practices for browse category and events configuration

The structure of your browse categories and corresponding parity in browse events directly impacts the model's ability to learn and optimize.

Balance granularity and traffic volume

This task requires:

  • Granular category structure: The browse model learns the click-and-purchase behavior associated (via user events) with each unique category string. If a category page, such as Electronics deals, gets substantial traffic, the model has a rich dataset to optimize rankings for that category. This helps the model accurately accumulate and interpret user behavior (clicks, purchases) for each category, leading to better ranking and personalization. Avoid generic handles like products which dilute the signal across different categories.

  • Optimized page category depth: Overly granular categories with low traffic volumes can lead to suboptimal model performance. With insufficient user interaction data, like clicks and add-to-carts, the model can't effectively give revenue-optimized rankings. Strike a balance between a detailed category taxonomy and ensuring that each category page generates enough traffic for meaningful model learning.

Ensure parity between API calls and user events

A critical technical requirement for successful model training is maintaining an exact match between the pageCategory string in your browse API calls and the corresponding browse user events. For example, if a user is browsing the Electronics deals category, the API call to retrieve browse results and the user events generated must use the identical category string, including the > delimiter.

For the training process, the model must join the API requests with the user events. A mismatch will also cause discrepancies in performance measurement. This alignment is monitored in the data quality dashboard. Failing to meet the threshold can keep your model from reaching higher tiers.

Monitor and troubleshoot recommendations

After setting up your website to get recommendations, we recommend that you set up alerts. See Set up an alert for prediction errors.

To troubleshoot errors, see Monitor and troubleshoot.