Kakao Reveals Ad Recommendation and Spam Filtering Technologies at KCC 2024

Kakao presented use cases of recommendation systems and spam filtering leveraging AI and ML technologies at the Kakao Tech Workshop during KCC 2024. The presentation was divided into two main topics: "Efforts for Better Recommendations" and "Efforts to Build a Clean Platform."

Enhancing Ad Recommendation Systems

Kakao introduced a bucket system for real-time ad recommendation serving. This system combines model synchronization and dynamic model decision-making methods, enabling the selection, modification, and management of models without changing or deploying serving code. Additionally, they shared scheduler configurations and operational practices for stably managing a large number of models optimized for various advertisers and objectives in batch inference environments.

For retargeting ads, they implemented fast data processing through a data-centric pipeline design considering reliability, responsiveness, and scalability. They also proposed measures to prevent performance degradation caused by overfitting through experiments to mitigate exposure bias and inductive bias (e.g., introducing exposure bias mitigation weights, changing MAB rewards).

LLM-Based Spam Response and Clean Platform

In the area of spam filtering, they covered cases applying AI technologies to both images and text. For adult image filtering, they resolved bottlenecks and improved performance by introducing Transformer models and improving the labeling process. For text spam response, they introduced LLMs to provide operators not only with classification results but also with "classification reasons," establishing a structure that assists in final judgments. Notably, they developed a proprietary spam LLM specialized for Kakao's regulatory policies, considering training data collection and hallucination prevention.