Target
visit propensity (predicted trip count) to frame and target offers
Target's TargetRun model predicts a Circle customer's trip count with Bi-LSTMs and feeds the CORE offer engine; an A/B test showed +78% redemption among customers who rarely redeem offers, +15% accuracy, and +40% predictive power.
Key points
- Prediction of Circle customers' visit frequency with the TargetRun model.
- Bi-LSTM deep learning and multi-head attention, on TensorFlow, PySpark, Kubeflow.
- A/B test: +78% redemption among customers who rarely use offers.
- Evidence B, confirmed status, TargetRun and the CORE engine in production.
Objective
Predict how many times a Target Circle customer will visit in the coming weeks in order to send them the right offers at the right time, and lift redemption among customers who rarely use offers.
The deployment
TargetRun is a deep learning model built by Target's MarTech Offer Personalization team. It predicts a customer's (Guest's) visit frequency over a two-to-three-week horizon. The architecture combines bidirectional LSTMs and multi-head attention over 52 weeks of customer history (spending, digital activity) summarized into sequences of 26 steps, plus static attributes (engagement, membership). The predicted number of visits feeds CORE (Contextual Offer Recommendation Engine), Target's offer engine based on a contextual bandit, which decides which offers a customer sees and how they are structured. CORE served millions of recommended offers in 2023.
Results Proof B
Official Target engineering blog describing the architecture and the results of a quantified A/B test, corroborated by a second official post on the CORE engine. Primary but internal source (no independent press), hence B.
How it works
Documented architectureThe stack in detail
- llm TargetRun (modele in-house) In-house deep learning model: bidirectional LSTMs plus multi-head attention over 52 weeks of history summarized into sequences of 26 steps, plus static attributes.
- plateforme CORE (Contextual Offer Recommendation Engine) In-house offer engine based on a contextual bandit, fed by the visit prediction; millions of offers served in 2023.
- outil TensorFlow Framework for training the sequential model.
- outil PySpark Distributed preparation of the customer history sequences.
- infra Kubeflow Orchestration of the ML pipelines (recurring training and scoring).
How it runs, concretely
For ops teams-
1Sequence construction data team / platform
52 weeks of spending and digital activity per customer are summarized into a sequence of 26 steps, plus static attributes.
-
2Prediction of the number of visits AI model (TargetRun)
The bidirectional LSTM model with attention predicts how many times the customer will visit in the coming weeks.
-
3Offer selection AI model (CORE)
The prediction feeds CORE (contextual bandit), which chooses the offers to show and their structure to maximize engagement.
-
4Delivery and measurement platform / data team
The offers go out in the Circle app and by email/push; the reaction (add, redemption) comes back as data and refines the models.
The predicted number of visits per customer. If a customer's digital and transactional history is too sparse, the prediction is noisy; the static attributes (engagement, membership) serve as an anchor.
How your customers perceive this type of use
Sourced studiesC'est la famille la moins acceptee : 68% des Americains jugent inacceptable un score financier personnel calcule par algorithme et 67% l'analyse video automatisee d'entretiens d'embauche (Pew Research, 2018). La demande d'explication et de recours est massive : 83% veulent savoir quelles donnees l'IA utilise et 91% veulent pouvoir corriger des donnees erronees (Consumer Reports, 2024). A l'echelle mondiale, seuls 46% se disent prets a faire confiance aux systemes d'IA et 70% jugent une regulation necessaire (KPMG / Universite de Melbourne, 2025).
Acceptance conditions
- Transparence sur les donnees utilisees : 83% des Americains la reclament (Consumer Reports 2024)
- Droit de correction des donnees erronees : 91% le demandent (Consumer Reports 2024)
- Explication de la logique de decision : 44% des consommateurs sont plus enclins a utiliser un agent IA si sa logique est clairement expliquee (Salesforce 2024)
- L'acceptabilite depend du contexte de la decision : 50% des Americains jugent equitable un score de risque criminel pour la liberation conditionnelle, contre 32% pour un score financier applique aux consommateurs (Pew Research 2018)
Red lines
- La decision opaque et sans recours sur l'emploi, le credit ou le logement : 45% tres mal a l'aise pour l'embauche, 39% pour le pret, 39% pour le logement (Consumer Reports 2024)
- Le scoring des personnes a partir de donnees comportementales : 68% le jugent inacceptable pour les offres financieres (Pew Research 2018)
Sources: Pew Research Center 2018 · Consumer Reports 2024 · KPMG / Universite de Melbourne 2025 · Salesforce 2024
How to replicate
Inference, not sourcedData prerequisites
- purchase and digital activity history per customer linked to a single ID
- loyalty program with offers and opt-in
- engagement signals (add, redemption)
Org prerequisites
- data science team able to maintain a sequential model
- downstream offer decision engine (bandit or recommendation)
Possible stack
- custom (TensorFlow / PySpark / Kubeflow)
- decisioning platforms (Braze, Salesforce, Adobe) with frequency prediction
The plan, step by step
- Step 1Frame the data: link 52 weeks of purchase and digital activity history to a single customer ID and build the sequences.Deliverable: Clean sequential dataset with static attributes (engagement, membership).
- Step 2Train a simple baseline then the sequential model (LSTM/Bi-LSTM) predicting the number of visits.Deliverable: Model beating the baseline on accuracy and over-prediction, validated out of sample.
- Step 3Connect the visit prediction to the offer engine (contextual bandit or rules) to frame targeting and timing.Deliverable: Offer targeting and timing driven by propensity, in pre-production.
- Step 4Run an A/B test on redemption, isolating the segment of customers who rarely use offers.Deliverable: Lift reading by segment and generalization decision.
First step: Predict a visit frequency (trip count) per customer over 2-3 weeks, then connect this prediction to the targeting and timing of the loyalty program's offers.
The same pattern, proven in other industries
Every Man Jack
predictive replenishment timing (predicted next order date)
Maersk
predictive monitoring of refrigerated cargo (customer visibility)
T-Mobile US
Intent-driven decisioning platform: an engine reads customer intent and sentiment in real time, decides the next action, and can execute it, beyond a simple chatbot
Sources
- S1 TargetRun: AI-Powered Trip Prediction for Delivering Timely Offers Primary archive pending
- S2 Contextual Offer Recommendations Engine at Target Primary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.