Automatic Playlist Continuation through a Composition of Collaborative Filters

TL;DR

This paper presents a collaborative filtering-based approach for automatic playlist continuation, combining multiple filters optimized with a TPE method, achieving competitive results in the RecSys Challenge 2018.

Contribution

It introduces a novel composition of collaborative filters optimized via TPE for playlist continuation, demonstrating effectiveness on a large real-world dataset.

Findings

Achieved 12th place out of 112 teams in the challenge

Effective combination of multiple collaborative filters

Demonstrated scalability on the Spotify MPD dataset

Abstract

The RecSys Challenge 2018 focused on automatic playlist continuation, i.e., the task was to recommend additional music tracks for playlists based on the playlist's title and/or a subset of the tracks that it already contains. The challenge is based on the Spotify Million Playlist Dataset (MPD), containing the tracks and the metadata from one million real-life playlists. This paper describes the automatic playlist continuation solution of team Latte, which is based on a composition of collaborative filters that each capture different aspects of a playlist, where the optimal combination of those collaborative filters is determined using a Tree-structured Parzen Estimator (TPE). The solution obtained the 12th place out of 112 participating teams in the final leaderboard. Team Latte participated in the main track of the challenge of the RecSys Challenge 2018.