Overview

A Warm-Up is a cross between a tutorial, an information-gathering phase, and a practice run. Unlike typical interfaces where diagrams can introduce the user to the experience, AI systems often require a small amount of user data to kick-start their personalization. Instead of shying away from this, design should lean in to this requirement, giving the user a special set of user interactions that helps them get acquainted with the system and explore its capabilities to find a style of interaction that suits them.

Theory

There exists a pervasive challenge with recommendation algorithms called the cold-start problem, which describes the precisely opposite phenomenon from the warm-up. In a cold-start scenario, the algorithm provides low-quality recommendations because it lacks any context or experience from past user interactions. When interactions between the user and these low-quality recommendations are measured (for use as future training data), the AI will often generalize incorrectly (e.g. assuming the user doesn’t want any clothing when the user rejectes a few recommended clothes). This can create a negative feedback spiral leading to both an awful user experience and bad data for further learning.

Instead, the warm-up attempts to create a stronger foundation for your system. You can provide the user with a few known examples to judge (that perhaps all users see initially), allowing the system to easily discern the user’s tastes as compared to others’. By leading the user through a few initial recommendations as well as the process or accepting or rejecting them, they are also trained on using a dynamic UI.

Examples

• Tell us what you’re into by Apple Music