Kakao Recommendation Team Unveils Personalized Recommendation Technology Combining Topic Modeling and MAB

The Kakao Recommendation Team introduced a personalized recommendation framework combining Topic Modeling and Multi-Armed Bandit (MAB). This approach applies topic modeling techniques from natural language processing to user-item click logs to compute topic vectors, providing real-time optimized recommendations through MABs assigned to each topic.

The key advantages of introducing MAB are as follows:

1. Learning Actual Objectives: It directly optimizes real-world KPIs such as actual click-through rates, rather than proxy objectives.
2. Exploration: It recommends new content that users may like but are not yet familiar with, considering long-term rewards.
3. Real-time Capability: User feedback is immediately reflected in the model, and experimental results confirmed that click-through rates significantly increase as the update loop becomes shorter.

The mechanism works by generating a final list through a weighted average of recommendation scores from each topic-specific MAB based on the user's topic vector, and updating the beta distribution of the corresponding topic-specific MAB when feedback occurs. This method offers the advantage of low computational cost and small model size while still providing personalized recommendations to a large number of users.