Kakao Arena 3rd Competition Reveals Melon's Music Recommendation Technology

Ahead of the Kakao Arena 3rd Competition (Melon Playlist Continuation), the core technologies of Melon's music recommendation system were revealed. This system performs a playlist generation task that generates similar songs and descriptive tags when a user inputs a mood keyword and a seed track.

Content-Based Filtering and Audio Processing

Melon processes audio data by converting it into mel-spectrogram format. This method reflects human auditory perception by adjusting the frequency axis to a logarithmic scale and grouping bands to preserve key information. The extracted spectrograms pass through convolutional models to be converted into embeddings, which are used to extract high-level features or recommend similar songs.

Collaborative Filtering and Autoencoder Usage

In the realm of collaborative filtering, Matrix Factorization and autoencoders are primarily used. Specifically, the denoising autoencoder technique is applied by representing playlists as 0s and 1s, hiding some tracks, and having the model reconstruct the hidden tracks using only the remaining ones. During this process, track lists and tags are learned together, modeling latent features to act as tags that describe the songs.

Embedding Operations and Tag Similarity Analysis

In the learned embedding space, meaningful results can be derived through vector operations. For example, adding the embeddings for the tags 'couple' and 'regret' yields tags related to greed or mistakes, while subtracting 'excitement' from 'love' yields tags related to religious love (agape). By leveraging this tag composability, customized playlists can be generated by combining seed tracks and tags.