Unique Research · Deep dive · January 7, 2025
In today’s fast-moving markets, knowing which customers to focus on is critical for firms that want to maximize returns. Traditional marketing often fails to keep up with shifting behavior, and opportunities slip away. To address that, researchers in the Katz School’s graduate Computer Science and Engineering division are turning to advanced tools such as reinforcement learning — a form of machine learning that mimics decision-making in dynamic, uncertain environments.
At the IEEE International Conference on Consumer Electronics in January 2025, Katz School researchers will present a paper titled “Optimizing Customer Targeting for Adaptive Marketing Strategies Using Reinforcement Learning and Neural Networks,” introducing a novel reinforcement-learning (RL) framework for customer targeting — a new way for enterprises to understand and respond to spending patterns.
“At the core of the framework are two breakthroughs: the mean strategy method and a neural-network approach based on the cross-entropy method,” said lead author Dr. David Li, director of the M.S. in Data Analytics and Visualization. “Together, these techniques let companies adjust strategy more effectively than ever.”
Reinforcement learning works by simulating interaction between an agent (here, the firm) and its environment (customers). The goal is to maximize return — finding and locking onto the most valuable customers. Each customer’s spending potential is modeled as a normal distribution, capturing both average spend and volatility.
The RL agent interacts with the environment over time, learning which customers may yield the highest return. The challenge is balancing exploration (trying new strategies to learn more) and exploitation (focusing on customers already identified as high-value). The mean strategy offers a pioneering way to strike that balance.
Classic RL methods such as ε-greedy and decaying ε-greedy randomly choose explore or exploit by a fixed probability. Useful as they are, they can miss opportunities — especially with diverse customer behavior.
The mean strategy improves this by using statistical confidence intervals to guide decisions dynamically. Instead of a fixed probability, it adjusts to how certain the model is about a customer’s spending habits. For example:
- If the model is highly confident a customer is high-value, it focuses on that customer (exploit).
- If uncertainty remains, it keeps exploring others so profitable opportunities are not missed.
“This data-driven approach yields more precise targeting while avoiding wasted effort on low-value customers,” said Zubair Khan, lead author and M.S. student in Artificial Intelligence.
Some behaviors are too complex for simple statistical models — that is where neural networks help. They excel at finding complex patterns in large data. Using the cross-entropy method, the network learns to fine-tune targeting by spotting hidden trends in spending.
For example, if high-value customers respond well to limited-time promotions, the network folds that pattern into strategy. Or consider a restaurant seeking maximum return from promotions: with the mean strategy it can offer an initial discount broadly, monitor spend, then shift focus to high spenders while still testing new segments — reaching the right audience without ignoring untapped opportunity. In e-commerce, RL plus neural nets helps identify high-margin customers best suited for personalized ads, discounts, or loyalty programs.
“Simulations with the framework show impressive results,” Li said. “Both the mean strategy and the neural approach outperform traditional methods such as ε-greedy at identifying high-value customers. They achieve higher cumulative returns and adapt better to changing behavior.”
The mean strategy’s ability to adjust exploration by statistical confidence is especially effective at balancing short-term gains with long-term learning. The cross-entropy method lets neural nets handle complex real-world scenes, giving firms a fine-grained, scalable targeting solution.
“Our work lays a solid foundation for the future of dynamic marketing,” Khan said. “By combining advanced statistics with machine learning, companies can stay ahead in understanding and adapting to customer behavior.”
— END —