# Cold-start Playlist Recommendation with Multitask Learning

**Authors:** Dawei Chen, Cheng Soon Ong, Aditya Krishna Menon  
**arXiv:** [1901.06125](https://arxiv.org/abs/1901.06125)  
**Date:** 2019-01-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1901.06125v1/

## Key points

- The paper explores playlist recommendation in three cold-start scenarios: cold playlists, cold users, and cold songs.
- The authors propose a multitask learning method to address these scenarios effectively.
- The method learns from user-curated playlists and uses a bipartite ranking loss to improve song recommendations within playlists.
- The authors demonstrate how the bipartite ranking loss can be approximated by minimizing a classification loss.
- Empirical experiments on real playlist datasets show that the proposed approach performs well in cold-start playlist recommendation tasks.
- The method generates personalized and relevant recommendations for existing and new users, as well as extends existing playlists with newly released songs.

## Summary

The paper titled "Cold-start Playlist Recommendation with Multitask Learning" by Dawei Chen, Cheng Soon Ong, and Aditya Krishna Menon explores the problem of playlist recommendation in three cold-start scenarios. The first scenario is cold playlists, where the goal is to recommend songs to create new personalized playlists for existing users. The second scenario is cold users, where the aim is to recommend songs to form new playlists for new users. Finally, the third scenario is cold songs, where the objective is to recommend newly released songs to extend existing playlists of users. To address these scenarios, the authors propose a flexible multitask learning method that can handle all three settings effectively. The proposed method learns from user-curated playlists and leverages a bipartite ranking loss to encourage songs within a playlist to be ranked higher than those that are not. The authors draw inspiration from an equivalence between bipartite ranking and binary classification and demonstrate how one can efficiently approximate an optimal solution of the multitask learning objective by minimizing a classification loss. The effectiveness of the proposed approach is evaluated through empirical experiments on two real playlist datasets. The results show that the method performs well in cold-start playlist recommendation tasks. In summary, this paper presents a novel multitask learning method for addressing cold-start playlist recommendation problems in various scenarios. The approach utilizes user-curated playlists and incorporates a bipartite ranking loss to improve song recommendations within playlists. Empirical evaluations demonstrate its effectiveness in generating personalized and relevant recommendations for both existing and new users as well as extending existing playlists with newly released songs.