Skip to content

PATTERN Cited by 1 source

Verbalization replaces feature engineering

Pattern

Replace the traditional recommender system stack of thousands of hand-crafted features with natural-language verbalizations of user histories, item metadata, and context — shifting the modeling discipline from feature engineering to context engineering.

Problem

Traditional recommender systems require: - Significant feature engineering for each new content type or product surface - Custom architectures for feature interactions and sequence modeling - Heavy feature infrastructure (feature stores, real-time pipelines, backfill jobs)

Adding a new content type (games, podcasts, live events) or surface requires substantial engineering investment across all these layers.

Solution

Verbalize raw interaction logs and metadata as text: - User engagement histories (plays, durations, ratings, list additions) - Item metadata (titles, genres, descriptions, cast) - Request context (device, surface, locale, time) - Task description (what the system should optimize for)

The LLM's pre-trained knowledge discovers cross-item relationships and evolving user interests that feature engineers would need to manually encode. New content types require only a new verbalization template — no feature pipeline changes.

Consequences

  • Low marginal cost for new content types or surfaces — just describe them in text.
  • Higher serving cost — natural language is more verbose than dense feature vectors.
  • Higher capability requirement — needs a strong LLM backbone with domain adaptation.
  • Data efficiency — GenRec matches a production ranker with 10–40× fewer labeled examples, suggesting the LLM's pre-training provides implicit feature engineering.
  • Context budget becomes the primary design constraint — replaces the feature-selection problem with a token-allocation problem.

Seen in

Last updated · 606 distilled / 1,847 read