Where Research Meets Real-world Business: The Inside Story of How We Successfully Deployed the Bandit Algorithm

――Congratulations on your paper being accepted at the WSDM 2025! Could you please give us an overview of this research?

Katsuragawa (PP): Thank you. Our research focused on applying a specific bandit algorithm, Top-Two Thompson Sampling (TTTS), to the problem of creative selection in online ads. In advertising, identifying the best creative, for instance, the one with the highest expected Click-Through Rate (CTR), is critical for maximizing an ad's effectiveness. We implemented TTTS in the Prism Partner DSP as a more efficient alternative to traditional A/B testing. This allows us to more accurately identify the best creative while reducing the costs of experimentation.

——To begin with, in what specific areas and for what challenges can the bandit algorithm be applied?

Ariu (AI Lab): I believe it can be used in almost any service that involves decision-making. Take advertising, our current use case: the algorithm is used to select the right ad creative by observing user reactions. You could imagine a similar application on ABEMA, choosing the best thumbnail or layout based on viewing history. Essentially, any scenario that requires sequential choice—making decisions based on past data—is a potential fit.

Abe (AI Lab): In the AI Division, we have already mastered the large-scale production of creatives with AI tools like our "KIWAMI-Prediction" series. The next challenge is figuring out 'how to effectively select and use' all those creatives. The bandit algorithm holds immense promise for solving this very problem. Our Reinforcement Learning (RL) team at AI Lab has been researching this area for some time, with numerous papers and real-world applications under our belt.

*What is the bandit algorithm? It is a method that learns to make optimal decisions while gathering data on its own. In online advertising, applications include identifying high-performing creatives.

――It sounds like the close collaboration between the business division and AI Lab was key to this success. Was there anything you consciously focused on to make that partnership work?

Abe (AI Lab): From the AI Lab side, a key principle was 'don't just push the latest methods.' As a researcher, it's tempting to use the latest technique for the sake of a paper, but we always start with the question: 'What's the best method for the product?' We run alongside the business teams with that mindset.

Ariu (AI Lab): Even though the TTTS method we used is published in a 2016 paper, it is somewhat dated given the research standards. There have been many follow-up studies. But when you consider physical operational constraints, like system limitations, a simpler implementation is often better. That’s why we proposed it.

Abe (AI Lab): The way we see it, AI Lab is constantly exposed to new research, so we build up a deep well of knowledge. We develop an intuition for which methods are likely to work in practice and which aren't. Our role is to then channel that knowledge back to the business divisions.

——So, by focusing on business value, you achieved success in both product improvement and paper acceptance. From the business side, what did you keep in mind when collaborating with AI Lab?

Kaneko (PP): The most critical factor was consulting with the AI Lab right from the planning stage. Early consulting gave us two advantages. First, they could suggest solutions we hadn't even considered, drawing on their deep research knowledge. Second, they could give us an early read on which approaches were unlikely to work, saving us from wasting time on dead ends. We also made a conscious effort to communicate frequently by proactively sharing product updates and setting up regular meetings.

Abe (AI Lab): People might think they need to have a perfectly defined problem before they come to AI Lab, but our team actually prefers to work together to clarify the fundamental goal—'What does the business or product really want to achieve?'—and then propose the best method. So we encourage people to come to us even at the 'I don't know if this is possible, but I have an idea' stage. No question or idea is too small.
 

――What's the appeal of an environment where research and business collaborate to create new value? Let's start with the two of you on the business side.

Katsuragawa (PP): It's incredibly empowering to work in where researchers not only understand business problems deeply but also help find solutions within real-world constraints. The best part is debating the optimal solution based on insights from cutting-edge research. And it's deeply rewarding to see the solution we built—one that is sometimes even recognized academically—directly fuel the business growth.

Kaneko (PP): As a data scientist in the field, a major goal was to achieve paper acceptance at top conferences like WSDM's Industry Day, The Web Conference, or KDD*, as proof of the quality of our outcome. AI Lab continuously produces world-class research, and the environment that allows for daily collaboration with them enabled us to elevate our practical work to a level recognized by the research community. Discussions with researchers from Google and Microsoft, and the high praise from the session chair, were very encouraging, and we are proud of our efforts in meeting global standards.

Working with AI Lab, I'm confident we can continue to aim for the world's top conferences. I'm excited to pursue new breakthroughs with others who share that same passion for technological exploration.

*The Web Conference and Knowledge Discovery and Data Mining (KDD) are top international conferences in the data mining field.