README.md

MelonRec-W2V

• Team ThunderCat: (Public LeaderBoard 58th)

1. Competition Overview

Goal: To predict songs and tags that were not given when half or all of the songs and tags in the playlist are unseen

data

• Playlist metadata (title, song, tag, number of likes, update time)
• Song metadata (song title, album title, genre, title date)
• Mel-spectrogram of the song

Number of participating teams: 786 teams

2. Issue

Challenge: all of the songs and tags in the playlist are unseen

Approach - Multi-Modal Retrieval - Query by Song, Tag, Title

Main Issue

1. Make co-embedding space that contain song vector, tag vector, title vector
2. Evaluate that the embedding space learns the semantic relationship.
3. Cover all retrieval scenario
   • Given Song to Song retrieval (Playlist Continuous)
   • Given Playlist Tag to Tag retrieval (Playlist Auto-tagging)
   • Given Playlist tag to Song retrieval (Unseen Item retrieval)
   • Title to Tag and Song retrieval (Sentence to Item retrieval)

3. Idea

Approach

• Make co-embedding space with Word2vec Method
• Single Modal Retrieval
  • Voting each modality
• Multi Modal Retrieval
  • Mean each modality
• Cluster Based Retrieval

4. Approach

1. Train Co-embedding Space (Word2Vec Embedding)

   • Input: Sentence (Title token, Tag, Genre, Song, Plylist id)
   • Ouput: Word, Item, Song, Plylist Vector

2. Multi-Modality Retrieval (Inference)

3. Evaluation with ndcg

   • tag wise, song wise

Dependencies

• numpy 1.16.2
• pandas 0.24.2
• matplotlib 3.0.3
• tqdm 4.31.1
• gensim 3.8.3
• sentencepiece 0.1.91
• sklearn 0.20.3
• khaiii
• pytorch 1.5.1

4. Folder and Files

• data download (link)
  • train.json, val.json, test.json, genre_gn_all.json, song_meta.json

Learning code (from scratch)

Best model hyper-parameter, window: 100, vector: 300, mincount: 10, iteration: 50, Skip-gram

$ train_embedding.py
$ embedding_most_similar.py

5. To-do

Measure Playlist-Song's Mean and Playlist's Vector Similarity

• Mid-Evaluation of Embedding Space: KL Divergence
• Training Method: Self-Supervised Approach

6. Reference

• Musical Word Embedding: Bridging the Gap between Listening Contexts and Music

  • Seungheon Doh, Jongpil Lee, Tae Hong Park, and Juhan Nam. Machine Learning for Media Discovery Workshop, International Conference on Machine Learning (ICML), 2020

• Automatic music playlist continuation via neighbor-based collaborative filtering and discriminative reweighting/reranking.

  • Zhu, L., He, B., Ji, M., Ju, C., & Chen, Y. (2018). In Proceedings of the ACM Recommender Systems Challenge 2018 (pp. 1-6).