New AI Turns Your Click History Into Smarter Recommendations
Your feeds and ads may get more relevant — while using less computing power.
Every time you open a shopping app, a video feed, or a news site, a computer quietly decides what to show you next. To do that well, it wants to know everything you've done there — not just today's clicks, but months or years of them. The problem: feeding thousands of past actions into a recommendation model is slow and expensive, so most systems either ignore old history or chop it off entirely. That's partly why recommendations can feel oddly forgetful.
A team of researchers has published a new approach called ChronicleRec. The idea is simple to describe: instead of re-reading your entire history every single time you load a page, the system compresses that history once into what it calls "Chronicle Tokens" — a compact, chronologically ordered summary of who you are and what you like. Recent activity is kept detailed, while distant history is blurred into broader patterns, like remembering last week vividly but recalling last year only in themes.
The clever part is timing. Older systems that summarize users often jumble the order of events or waste space repeating themselves. ChronicleRec anchors each summary to a moment in time, so when the system asks "what should we show this person now?" it only looks at history that happened before that point. The authors trained it to predict your recent behavior from your older history, effectively teaching it to line up long-term patterns with what you want today. Because these summaries don't depend on the specific item being recommended, they can be saved and reused.
The results are promising. Tests on two large public datasets, plus a seven-day live experiment on real advertising traffic, showed the method beating simpler alternatives and getting close to the accuracy of vastly more expensive full-history models. For ordinary users, the practical effect is recommendations that remember your long-term interests instead of fixating on whatever you clicked five minutes ago — and systems that cost companies less to run.
- ChronicleRec compresses years of your clicks into one reusable summary, so apps don't re-read your whole history every time.
- It keeps recent behavior detailed while blurring old behavior into patterns, much like human memory works.
- A seven-day live ad test showed real-world gains, meaning better-targeted feeds and ads — and lower computing costs.
Why It Matters
Better recommendations mean less time scrolling past junk, but also more precise tracking of your habits.