Ecommerce ops · Production

DoorDash personalizes store feed using store2vec vector embeddings and gradient-boosted models

The problem

Geographic constraints caused severe sparsity in DoorDash's consumer-to-store matrix, making standard collaborative filtering impractical; the initial knowledge-based recommender lacked latent semantic signals about store similarity and consumer preferences.

Workflow diagram · grounded in source
1
Session order data collected
trigger
“each sentence is a list of stores viewed together in a user session”
2
Store vector embedding
ai_action
“we embed stores as vectors using the word2vec (CBOW) algorithm from gensim package with the following modification. - each store is a word in our vocabulary and - each sentence is a list of stores viewed together in a user session. For w…”
3
Consumer vector generation
ai_action
“to generate vectors for a consumer, we sum the vectors for each store they ordered from in the past 6 months or 100 orders. To then determine the distance between a store and a consumer, we take the cosine distance between the store's ve…”
4
Feature generation
ai_action
“we extract many features having to do with the annotated data on consumer and stores such as categories, rating, popularity, and browse / click / order information”
5
Model training
ai_action
“We train logistic regression and gradient-boosted machine (GBM) models. For GBM models, we use LightGBM. These are the same frameworks we use for many other machine learning systems at DoorDash”
6
Model evaluation
validation
“The model is predicting P(order | consumer, store) and is a binary classifier. To evaluate it for this ranking problem, we use area under curve (AUC) of the precision/recall curve”
7
Personalized recommendations served
output
“are in the process of testing and rolling out these changes more broadly in email and in-app”
Reported outcome

By incorporating store2vec latent features and gradient-boosted machine models, DoorDash saw approximately 20% improvement in offline AUC and approximately 5% increase in click-through rate in initial email tests, building on a prior 25% CTR gain from the initial recommendations system.

Reported metrics
Click-through rate increase vs popularity baseline25%
additional CTR increase in initial email tests5%
offline AUC improvement20%
Reported stack
store2vecword2vecgensimLightGBM
Source
https://careersatdoordash.com/blog/personalized-store-feed-with-vector-embeddings/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

By incorporating store2vec latent features and gradient-boosted machine models, DoorDash saw approximately 20% improvement in offline AUC and approximately 5% increase in click-through rate in initial email tests, bui…

What tools did this team use?

store2vec, word2vec, gensim, LightGBM.

What results were reported?

Click-through rate increase vs popularity baseline: 25%; additional CTR increase in initial email tests: 5%; offline AUC improvement: 20% (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Session order data collected → Store vector embedding → Consumer vector generation → Feature generation → Model training → Model evaluation → Personalized recommendations served.

Related ecommerce ops cases