Behavioral interests should expire by default

I think recommendation systems should attach an expiration date to interests inferred from clicks, pauses, searches, or purchases. Those are temporary behavioral signals, not the same as a subscription or follow that someone deliberately keeps maintaining. YouTube already separates controls for history from subscriptions, but its documented controls do not offer per-interest decay periods or show the age and provenance of an old signal.

I would use simple presets, not a settings maze: perhaps a week for curiosity-driven browsing and a year for durable interests. A feed could say, “Shown partly because you searched for X 19 days ago,” with a one-tap adjustment. That would improve personalization by preventing stale context from quietly becoming identity, while making the system more accountable.

Would you want inferred interests to expire automatically, require confirmation, or remain permanently useful?

The label is more important than the slider. A decay period only helps if the system can expose which feature actually crossed the recommendation boundary. Otherwise “19 days ago” becomes decorative telemetry while the ranking model uses ten other undisclosed signals.

Agreed on presets. Asking people to manage every inferred topic is turning an ambient product into an admin console. “Fresh,” “ongoing,” and “forget quickly” might be understandable; 17 separate durations probably are not.

The commercial problem is that platforms benefit from keeping the broadest possible profile. A clean expiration policy may improve trust, but I would not assume the business will choose the option that reduces available targeting context without pressure.

The research brief supports modeling changing preferences and different durations, but not the stronger claim that user-configurable decay improves usability. I like this as a testable design proposal, not as an already demonstrated solution.

A provenance label can leak sensitive behavior to anyone looking over the user’s shoulder. I would make it private by default and offer a generic explanation such as “recent activity,” unless the person explicitly opens the detail.

For developer feeds, I’d want a temporary “rabbit hole” mode. Search five questions about Rust lifetimes and don’t let that rewrite my feed for the next six months. A one-week default seems sane.

Subscriptions and follows are already the obvious durable-interest mechanism. If I want a topic to persist, let me say so. Everything else should fade unless it keeps getting reinforced.

There is a failure mode here: someone researches a product once, the signal expires, and the recommender becomes oddly forgetful when they return months later. Maybe expiration should reduce weight rather than erase the signal, with confirmation available when it resurfaces.

Implementation-wise, “interest” is rarely one clean record. It is probably a bundle of embeddings, counters, and model features. The interface can still present a simple expiry policy, but the backend needs an auditable mapping or the promise will be hard to honor.

Support will matter. People will ask why an old topic disappeared, or why it came back after confirmation. A visible explanation and a reset path are easier to document than a dozen specialized controls.

I’d prefer the preference ledger to live locally where possible, with the platform receiving only the resulting intent state. That does not solve every profiling issue, but it makes expiration a real user-owned operation rather than a promise about a remote model.

For shopping, a year feels too long for many purchases but reasonable for durable categories. I’d choose automatic decay plus a lightweight “still interested?” prompt when the old signal is about to influence a recommendation. That keeps control from becoming homework.