How to Train Your Algorithm
Go through the master “red team” lists with the class, so that everyone knows all the issues that were identified with each optimization goal.
Ask if they can think of any examples where an app or website’s optimization goals backfired or led to bad outcomes. You can prompt them with these examples:
- YouTube optimizing for watch time led to videos getting longer and longer
- Snapchat optimizing for daily use made people spend an hour every day sending blank pictures just to keep up their Streaks
- Different sites optimizing for virality made “fake news” stories spread more widely than true ones
- Other sites optimizing for engagement made extreme, offensive and harassing posts get recommended over more civil ones.
Then ask:
Think back to the video app you analyzed. Now that you know some of the concerns with the different things an app can be optimized for, do you agree with its optimization goals were ranked?
Why or why not?
Next, have students brainstorm other optimization goals that algorithms could be aiming for. You can prompt students with these examples:
- Accuracy
- Civility (promoting positive interactions between users)
- Diversity (not showing you too much of the same thing)
After students have discussed these questions for a few minutes, explain to students that while we can’t always affect what algorithms are optimized for, we can often train our algorithms to show us more of what we want, and less of what we don’t want.
Show the video How to Train Your Algorithm in fullscreen or have students access it, and then answer the questions in the student chapter How to Train Your Algorithm.
Select this link to access a description for people who are blind or low-vision.
Training an algorithm to achieve a particular goal.