Can language models discover what users actually want from activity logs?

Recommender systems predict the next item a user might click on, given their history. But when you ask users what they're actually doing on the platform, they describe something different: persistent, overarching interests — "designing hydroponic systems for small spaces," "learning the ukulele as a beginner," "cooking Italian recipes." These are interest journeys, and they operate at a completely different level of abstraction from next-item prediction.

Survey data shows 66% of respondents recently pursued a valued journey on the platform. Of those, 80% consumed relevant content for more than a month, with half saying some journeys last more than a year. People pursue 1-3 journeys simultaneously.

The semantic gap is real: collaborative filtering captures correlational patterns between items ("people who watched X also watched Y") but cannot reason about the user's underlying goal, need, or interest. Two users both interested in stand-up comedy may pursue completely different aspects — history documentaries vs. SNL skits. The journey is personalized at a granularity collaborative filtering cannot reach.

LLMs can bridge this gap. Through personalized clustering of user activity logs followed by LLM-powered journey naming, the system produces journey descriptions users identify with. But specificity matters — "greenhouse designs for cold climates" was irrelevant for someone pursuing indoor gardening. The right level of abstraction is what the user would actually say to a friend asking about their interests.

This connects to How do personalization granularity levels trade precision against scalability? — interest journeys operate at the user level but require persona-level precision. Since Does chatbot personalization build trust or expose privacy risks?, journey-aware systems that understand your persistent interests will trigger both the trust and privacy dimensions of this dual dynamic.