The Fifth Elephant
The Fifth Elephant

The Fifth Elephant 2026 Winter Edition

Call for Data Problems and Proposals

All submissions

Abhishek Anand

Beyond Prediction: Adaptive Audience Selection with Bayesian Bandits and Constrained Optimization

Submitted Oct 11, 2026

🎤 I am submitting: To speak I have a submission for: Showcase - a cool idea I have built/working on - 20 mins

Submission type - Showcase - new ideas - 20 mins

Describe your session in 2 paragraphs

In digital advertising, selecting the right audience is not a one-time decision. Campaigns evolve continuously as new impressions are delivered, conversions are observed, and audience availability changes. Yet audience targeting often relies on segments selected before a campaign begins, limiting the ability to respond to emerging behavioral signals. Optimizing these decisions introduces several challenges: delivery feedback is incomplete, conversions are sparse, and selecting high-performing audiences alone does not guarantee that a campaign can meet its delivery goals. How do we build a system that learns from limited feedback, balances exploration with exploitation, and makes better targeting decisions without compromising operational constraints?

In this showcase, we present an adaptive audience recommendation framework developed at Roundel that combines Positive-Unlabeled (PU) learning, conversion propensity modeling, hierarchical Bayesian multi-armed bandit simulations, and constraint-aware optimization. We first model the likelihood of ad delivery and subsequent conversion, addressing the absence of explicit negative delivery labels through PU learning. These calibrated probabilities feed a hierarchical Bayesian bandit simulation, where audience segments act as arms and delivery and conversion are modeled as sequential outcomes. Finally, an optimization layer selects audience segments that maximize predicted conversion potential while satisfying inventory and delivery constraints. We will walk through the mathematical formulation and design of the framework and the practical challenges of scaling predictions to millions of users across live campaigns. The session highlights how predictive ML, uncertainty-aware simulation, and mathematical optimization can be combined into a practical decision-making system.

Mention 1–2 takeaways from your session

  • Moving from predictions to decisions: Learning how calibrated propensity models, Bayesian bandit simulations, and constrained optimization can turn user-level predictions into actionable audience recommendations while balancing exploration, exploitation, and operational feasibility.
  • Learning from imperfect feedback: Exploring how positive–unlabeled delivery data and sequential delivery-to-conversion outcomes can be modeled and evaluated at scale, even when clean experimental comparisons are difficult.

Which audiences will benefit most from your session?

Data scientists, ML engineers, and applied AI practitioners working on recommendation systems, sequential decision making, and optimization under uncertainty. The session will also benefit professionals in advertising technology and retail media interested in adaptive audience targeting and campaign optimization.

Add your bio

Abhishek Anand is a Lead Data Scientist at Target, working with Roundel, Target’s retail media business, on predictive modeling, advertising optimization and audience recommendation systems. His expertise spans predictive modeling, Bayesian bandits, constrained optimization, and agentic AI applications for media planning.

Topic/domain tags
Applied Machine Learning, Decision Optimization, Advertising Technology (AdTech)

All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures