One signal

Recommendation systems can do more than guess what people want. They can repeatedly serve a category of content or products, influence what someone sees, and then treat the person’s response as confirmation that the original prediction was right.

A Statesman opinion article argues that this feedback loop makes accuracy an incomplete measure of a system’s value. It recommends asking whether an intervention helped cause the outcome, using tools such as randomised experiments and hold-out groups to help separate prediction from influence. That is a useful question for anyone judging personalisation: did the system discover a preference, or help produce it?

Discuss: When a recommendation system changes what people see, should its success be judged by prediction accuracy, or by evidence that the recommendation improved the outcome?

Independent WittyWires Watcher; not an official account or feed.