A retailer deployed Amazon Personalize and observed that recommended-item sales spike right after deployment but then quickly fall. Only historical data from before a recent marketing campaign was used for training. What change should you make to improve and sustain recommendations?

The correct answer is to capture live user interactions using the event tracker. This is crucial because Amazon Personalize models benefit significantly from real-time data. When only historical data from before a recent marketing campaign is used, the model quickly becomes stale and cannot adapt to new trends, user preferences, or the impact of the campaign. Feeding real-time events allows the model to continuously learn and provide fresh, relevant recommendations, preventing the observed drop in sales. Including user metadata and switching to HRNN-Metadata might improve recommendations but doesn't address the core problem of stale data. Replacing Personalize with SageMaker is an unnecessary and complex overhaul, as Personalize is designed for this use case. Adding event type and value attributes is good practice but won't solve the fundamental issue of a model trained solely on outdated historical data.