Kakao Improves Valid Play Rate by 20% with MoE-Based Personalized Ranking Model Using GDCN

Kakao developed a deep learning-based personalized ranking model combining GDCN (Gated DCN) and MoE (Mixture of Experts) to overcome the limitations of existing segment-based recommendations. This model improved performance across all user segments, from new to heavy users, through RLTW labeling that mitigates watch time bias and an expert branching structure based on user activity.

Architecture and Model Design

The system consists of a feature pipeline, training pipeline, GPU inference server, and serving API server. During inference, only modified users/items are incrementally served to minimize write load, scoring 300 candidates. The model evolved to GDCN to address the limitations of DCNv2, applying an Information Gate to Low-rank cross interactions (rank=32) to select feature reflection weights. Additionally, a Single-gate 3-Experts MoE structure optimizes predictive power for light and heavy users respectively.

Performance Improvements and Operational Results

In offline experiments, the GDCN+MoE model recorded improvements of +35.0% in average watch time and +6.5% in watch ratio. Increasing the number of candidates to 900 resulted in only a +0.9%p improvement in VTR, deemed inefficient relative to latency costs, leading to the final decision to operate with 300 candidates. After actual deployment, the Valid Play UV metric rose in a stepwise manner, increasing by +10%p (relative +20%), and recommendation diversity expanded by +53.7%.