ecommerce_ops · workflow
DoorDash builds a universal ranker with UCB exploration for homepage personalization
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.
How it works
Common implementation structure
How this type of workflow is generally built, generalized across documented cases — not tied to any one vendor's stack. Click any stage to read what happens there. Specific products that implement these stages appear in “Tools commonly seen” below.
Stage 1 · FPR candidate retrieval
First pass ranking selects no more than 1,200 candidates from ElasticSearch most relevant to the consumer.
Tools used
ElasticSearchPyTorchFabricatorLSTM
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.
What failed first
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.
Grounding & classification
Source type: technical build writeup
19 fields verified against source quotes.
personalizationpredictive analyticsrecommendation systemproduct catalogbuilder submittedmetric backednamed customerproduction runtime claimedtools describedworkflow describedecommerceconversion increasetechnical build writeupecommerce ops