What level of abstraction makes interest journeys feel personally relevant to users?
This explores the tension between abstraction and specificity in personalization — at what 'altitude' a description of someone's evolving interests lands as recognizably theirs rather than generic or merely a log of clicks.
This explores the tension between abstraction and specificity in personalization: not too granular to be noise, not so generalized that it stops feeling like *yours*. The corpus is interesting here because it pulls in two directions at once, and the resolution turns out to be a specific kind of middle altitude.
On one side, abstraction wins. The PRIME work shows that semantic memory — distilled preference summaries — consistently beats episodic memory, the literal replay of your past interactions Does abstract preference knowledge outperform specific interaction recall?. Storing 'what this person tends to care about' outperforms storing 'here are 40 things they clicked.' So pure specificity, the raw activity log, is the wrong altitude — it's the thing systems should abstract *away from*.
But climb too high and personal relevance evaporates. There's a quiet structural pull toward over-abstraction: common words express more general meanings, and LLMs are biased toward common words, so paraphrasing systematically drifts upward into hypernyms and erases the expert-level specificity that made a description distinctive Does word frequency correlate with semantic abstraction?. 'Interested in gardening' is true of millions; it recognizes no one. The relevance lives in the rung below.
The interest-journeys work names that rung precisely. The win wasn't an abstract category but a phrase like *'designing hydroponic systems for small spaces'* — described at user-level granularity with persona-level precision, capturing journeys that last over a month and that collaborative filtering misses entirely Can language models discover what users actually want from activity logs?. That's the answer to the altitude question: high enough to be a coherent pursuit rather than a click, low enough to be unmistakably one person's. It's a *named project*, not a *taxonomy node*.
Two lateral notes sharpen this. First, a single altitude per person is already wrong — people hold multiple personas, and the relevant abstraction is the one tied to the specific thing being recommended, weighted dynamically rather than averaged into one blurry vector Can attention mechanisms reveal which user taste explains each recommendation?. Second, even at the right altitude, accuracy-optimized ranking quietly collapses a person's secondary interests into their dominant one, so calibration matters as much as abstraction level — proportional representation of *which* journeys, not just how they're phrased Do accuracy-optimized recommendations preserve user interest diversity?. The thing that feels personally relevant, then, isn't a single well-pitched summary — it's a small set of named, durable pursuits, kept in their real proportions.
Sources 5 notes
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
WordNet analysis shows hypernyms (general concepts) occur more frequently than hyponyms (specific ones). Combined with LLMs' frequency bias, this means preferring common paraphrases systematically drifts toward abstraction, erasing expert-level specificity.
66% of users pursue valued interest journeys lasting over a month, described in specific phrases like 'designing hydroponic systems for small spaces.' LLM-powered journey discovery bridges the semantic gap that collaborative filtering cannot reach, operating at user-level granularity with persona-level precision.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.
Steck's research shows that ranking by per-item relevance naturally produces lists dominated by a user's primary interest, even when they have documented secondary interests. Enforcing calibration via post-hoc reranking restores proportional representation without sacrificing overall accuracy.