Ecommerce ops · Production

DoorDash builds a universal ranker with UCB exploration for homepage personalization

The problem

DoorDash's homepage used a fixed entity ordering (carousels first, stores second) that worked initially but became suboptimal as the carousel count grew dramatically, burying individual stores. Pure exploitation of known preferences also risked filter bubbles and unfair treatment of newer merchants.

First attempt

A fixed ordering of carousels before stores stopped scaling as carousel count exploded. Building a dedicated ranking model per entity type was impractical, and calibrating scores across mixed entity types proved nontrivial.

Workflow diagram · grounded in source
1
FPR candidate retrieval
integration
“Selecting no more than 1,200 candidates from ElasticSearch that are most relevant to the consumer experience among all stores”
2
SPR filtering and pre-ranking
ai_action
“Second pass ranking, or SPR, filters and pre-ranks those candidates, ultimately: Choosing up to 50 for the first page of the store feed”
3
Universal Ranker pConv scoring
ai_action
“The UR jointly ranks vertical positions for mixed types of entities by order of their pConv — probability of conversion, which is how we measure the relevance of a recommendation”
4
UCB uncertainty blending
ai_action
“the final UCB score is then obtained by blending the UR score (pConv) with the uncertainty”
5
Final ranked homepage output
output
“Final ranking, or FR, is the concluding stage, resulting in vertical rankings of all available carousels and stores on the first page of the store feed”
6
Daily engagement data refresh
feedback_loop
“Data driving the UR and uncertainty is refreshed daily”
Reported outcome

The combined Universal Ranker and UCB framework showed consistent improvements in online experiments, driving more consumer engagement and conversions on the homepage and increasing consumer interest in new merchants.

Reported metrics
Homepage recommendation performanceconsistent improvements in our online experiments
Consumer homepage engagement and conversionmore consumers to engage with our recommendations and convert on the homepage
Consumer exploration of new merchantsMore consumers are interested in trying new merchants and items they never ordered before
Reported stack
ElasticSearchPyTorchFabricatorLSTM
Source
https://careersatdoordash.com/blog/homepage-recommendation-with-exploitation-and-exploration/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The combined Universal Ranker and UCB framework showed consistent improvements in online experiments, driving more consumer engagement and conversions on the homepage and increasing consumer interest in new merchants.

What tools did this team use?

ElasticSearch, PyTorch, Fabricator, LSTM.

What results were reported?

Homepage recommendation performance: consistent improvements in our online experiments; Consumer homepage engagement and conversion: more consumers to engage with our recommendations and convert on the homepage; Consumer exploration of new merchants: More consumers are interested in trying new merchants and items they never ordered before (source-reported, not independently verified).

What failed first in this deployment?

A fixed ordering of carousels before stores stopped scaling as carousel count exploded.

How is this ecommerce ops AI workflow structured?

FPR candidate retrieval → SPR filtering and pre-ranking → Universal Ranker pConv scoring → UCB uncertainty blending → Final ranked homepage output → Daily engagement data refresh.

Related ecommerce ops cases