Consumer embeddings improve ad ranking for food delivery marketplaces
PALATE: Pretrained Autoregressive LATent Embeddings
Information RetrievalMachine Learning
Summary
Personalizing ads for consumers on local-commerce platforms is hard because each person interacts with only a few stores, making their history sparse. The authors built a model that predicts what stores a person might engage with next by learning from their past clicks and purchases. This model produces a single summary vector representing a person's preferences, which can then be used to show better ads across many surfaces. Their approach improved ad ranking accuracy offline and boosted revenue and conversions in live tests. Key innovations include training the model to focus more on new businesses, a two-step training process, and a carefully chosen set of negative examples.
What this means in practice
- •For e-commerce engineers: Create consumer preference embeddings from interaction sequences to improve ad ranking and personalization on marketplace platforms.
- •For digital marketing teams: Increase revenue and conversion by integrating single consumer embeddings into ad ranking and targeting systems.$Commercial implications: Enables improved personalized advertising products that increase marketplace monetization.
Authors
Sizhang Zhao, Praveen Gondhi, Bin Li, Utsaw Kumar
Abstract
Local-commerce and food-delivery marketplaces run large ads businesses: across hundreds of thousands of stores, merchants pay to be surfaced on browse and search surfaces, and a ranker decides which ads each consumer sees. Ranking well demands strong per-consumer personalization, yet a consumer's raw interaction history is sparse and high-cardinality. Our production ranker had personalized only through hand-engineered signals: aggregate engagement counts, categorical embeddings, and order-agnostic pooled history, with no learned per-consumer sequence representation. We learn one by training a GPT-style generative model of the consumer's engagement sequence to predict their next few engagements, and read its last-token hidden state as the consumer embedding. That single frozen vector becomes one shared personalization input for every ad surface: a daily job refreshes it, and each ranker consumes it as an ordinary dense feature. Offline it is the largest single-feature gain in our ranking-AUC ladder, a relative Eval-AUC gain of +0.59% on a browse-driven feed and +0.57% on a query-driven search surface. Online, it shipped as one component of bundled ranker retrains on both surfaces: Discovery (revenue +0.79%, conversion rate +1.17%) and Search (revenue +0.94%, conversion rate +1.50%). Three design choices adapt the generative objective to a repurchase-heavy marketplace, and are the core of our contribution: an exploration-weighted loss that up-weights businesses new to the consumer over ones already engaged; a two-stage curriculum: a click pretrain that keeps the representation broad and exploratory, then a conversion fine-tune that sharpens purchase intent; and a designed negative set that pairs a fixed popular-business pool and businesses shown but not ordered with random and in-batch negatives.