Research Paper

Core Innovation

This paper introduces BanditLP, which unifies neural Thompson Sampling for learning objective-specific outcomes with a large-scale linear program for constrained action selection. Unlike prior work, it supports arbitrary neural architectures and scales to billions of variables, enabling practical deployment of constrained contextual bandits at web scale.

Why It Matters

Personalized recommendation systems must balance multiple objectives and constraints at scale, which is challenging with existing methods. BanditLP improves decision quality by integrating exploration and constrained optimization, leading to better user engagement and business metrics. Its scalability enables deployment in large-scale production environments, transforming recommendation workflows.

Potential Customers & Pain Points

• E-commerce platforms – Need scalable personalized recommendations under business constraints
• Digital marketing firms – Require optimized multi-objective campaign targeting
• Streaming services – Must balance content diversity and user preferences at scale
• Ad tech companies – Need efficient constrained bidding and targeting strategies.

Business Model

SaaS platform licensing with tiered pricing based on scale and features; enterprise consulting and integration services for large customers.

Competitive Landscape

• Google Ads
• Amazon Personalize
• Microsoft Azure Personalizer
• Criteo
• Taboola

Implementation Challenges

• Integration complexity with existing recommendation systems
• Computational resource requirements for large-scale LP solving
• Adoption resistance due to operational changes in marketing workflows

Validation Strategy

• Pilot deployments with major digital marketing and e-commerce clients
• Benchmarking against existing recommendation and bandit algorithms on public datasets
• Measuring business KPIs such as engagement
• conversion
• and revenue lift in production