LCM-Promo: An LLM-Based Uplift Model for Promotion Allocation

LCM-Promo. An LLM-Based Uplift Model for Promotion Allocation. By Nesta Midavaine and six others, from Prosus and iFood. Published on the twenty-seventh of October.

Promotion allocation is the question of which users should receive which promotions. The tool proposed for it is an uplift model, and the outcomes that matter in the end are two: whether users convert, and what the promotions cost. A model for this task has to understand the user well enough to predict both intent and response to an offer.

LCM-Promo approaches the problem with a language model. It is described as a task-specific Large Commerce Model, or LCM: a large language model finetuned on transactional commerce data, trained to predict user intent and how a user responds to a promotion. The uplift model, in other words, is a language model that has been taught the language of transactions.

Its outputs are read in a particular way. The model learns promotion-response curves from production logs, and those curves are monotonic. To get them, it interprets single-token outputs as probabilities of crossing a conversion threshold. It also conditions on the context of the session, such as the day and the delivery shift, so the same user can be treated differently depending on when they arrive.

The test was a production A/B test on churn reactivation. Relative to the Segment-EMA production baseline, LCM-Promo increased conversion by one percent and reduced promotion cost by six percent.

Both of the outcomes named at the start moved in the favourable direction at once, and they did so in production rather than in an offline evaluation. Slightly more users converted, and the promotions spent on them cost less. The ingredients behind that are few: a language model finetuned on commerce data, single-token outputs read as threshold probabilities, and the context of the session.