New AI Learns When You're Unsure to Fix Bad Recommendations
Fewer useless suggestions and faster choices — this could change how shopping apps guess what you want.
Recommendation engines — the software behind "you might also like" — mostly assume you have stable, clear preferences. A team publishing at WISE 2026, a data-mining conference, says that's wrong. Sometimes you know exactly what you want; sometimes you're just browsing. Their system, called UHIFlow, measures how uncertain it is about your intent, then changes how it behaves based on that.
In plain terms, it reads two kinds of clues: pictures (product photos, thumbnails) and text (titles, descriptions, reviews), plus your behavior. A technique called "flow matching" helps it track how those clues gradually shift instead of snapping to one fixed answer. When the clues conflict or are weak, the system flags high uncertainty.
That flag changes the response. With high uncertainty, it offers broader categories — "here are some cozy sweaters" rather than "this exact navy size medium." When your signals are clear, it gets specific. The researchers tested it on three real-world datasets and report it outperformed existing methods.
The catch: this is a conference paper, not a product. No app uses it, and results on datasets don't always survive contact with real users. There's also a privacy angle — a system that reads your hesitation is still a system tracking you more closely. Still, it points toward recommendations that waste less of your time.
- Most 'you might also like' engines assume your taste never wavers — this one notices when you're actually undecided.
- It combines picture and text clues, then gives broad suggestions when signals are murky and specific ones when they're clear.
- Tested on three real datasets and accepted at the WISE 2026 conference, but no app uses it yet.
Why It Matters
Smarter recommendations could save you scrolling time and cut the frustration of useless online suggestions.