Our Solution to the Kakao Arena - Melon Playlist Continuation Challenge (2nd Place on the Final Leaderboard)

This is our second-place solution for the Kakao Arena - Melon Playlist Continuation Challenge (Song nDCG: 0.321238, Tag nDCG: 0.507595). The task is to build a model that recommends songs relevant to playlists.

Enviroment

• Python 3.7.4
• Intel Core i9-9960X+RAM 120GB Machine
• The preprocessing step took 10-20 minitues, and the training and inference step took 2-3 hours (used 40GB memory).

Dependencies

• numpy/scipy
• scikit-learn (0.22.2.post1)

Model

• This is a Neighbor-based CF model. The model estimates the similarity between songs included in a playlist u (item i) and songs not included (item j) as follows, and the most similar songs is recommended.

• We used a discriminative reweighting technique to improve model performance by penalizing songs that do not represent their playlists (Zhu et al., 2018). For example, if some items in a playlist are unrelated to the other songs in the playlist, we apply panelty to these items, making them less likely to be recommended. To accomplish this, we generate a vector r_j for each item j that represents the similarity between each item j and the entire set of songs in the target playlist. Then, using r_j, which learns the playlist representativeness of each song j, we train an L2-regularized support vector classifier (SVC) that predicts y_j (1 if a playlist u contains an item j, 0 otherwise).

• We use title-related keywords that appear 5 times or more in the dataset to reduce the cold start issue. Keywords are included in playlist items alongside other songs in the playlist.

Dataset

Download train.json, val.json, test.json into res/ folder (https://arena.kakao.com/).

How to run

1. Run python preprocess.py to preprocess the dataset.
2. Run python inference.py to generate recommendations which are saved in results.json.

Reference

Zhu, L., He, B., Ji, M., Ju, C., & Chen, Y. (2018). Automatic music playlist continuation via neighbor-based collaborative filtering and discriminative reweighting/reranking. In Proceedings of the ACM Recommender Systems Challenge 2018 (pp. 1-6). https://github.com/LauraBowenHe/Recsys-Spotify-2018-challenge