Kakao Reveals Related Recommendation Technology Used in KakaoPage and Melon

Kakao's recommendation team revealed the principles and effectiveness of the related recommendation technology utilized in KakaoPage and Melon. Related recommendation is a method that expands user interests by suggesting content with similar outcomes after a user consumes specific content.

KakaoPage: Text-Based Vectorization and Real-Time Optimization

KakaoPage provides related recommendations on work introduction pages and at the end of the final episode. It determines similarity by converting the work's title and introduction text into points in a vector space using Word2Vec or TF-IDF weights. Subsequently, it measures the expected value of target metrics (such as click-through rate) based on Thompson Sampling and adjusts recommendation probabilities in real time according to user actions after exposure.

Melon: Similar Song Recommendation Techniques and Effects

Melon's similar song recommendations are generated by combining various techniques.

• Collaborative Filtering (CF): Analyzes users' song consumption patterns to measure similarity between vectors.
• Matrix Factorization: Extracts item vectors using ALS or BPR techniques, or deep learning models like Auto-Encoder.
• Item2Vec: Identifies the context of songs within playlists to recommend songs with similar concepts.
• Audio Data Analysis: Converts audio into mel-spectrograms and trains deep learning models to define similarity between audio tracks, complementing the limitations of CF (such as cold start).
• Ranking Correction: Ensures diversity by correcting the predicted preference scores of recommendation models based on streaming conversion rates.

Performance Metrics

Users who used similar song recommendations on Melon had approximately 24% higher weekly average streaming numbers compared to non-users, and the average number of unique songs streamed per person was 53% higher. This demonstrates that related recommendations are effective in helping users explore and consume new content.