MMCF: Multimodal Collaborative Filtering for Automatic Playlist Continuation

Automatic playlist continuation (APC) is a common task of music recommender systems, enabling the automatic discovery of tracks that fit into a given playlist. To recommend a coherent list of tracks to users, it is important to capture the underlying characteristics of a playlist. Unfortunately, existing recommender models suffer from several problems: (1) They tend to misinterpret tracks that appear rarely in a playlist (popularity bias) (2) they cannot extend user's playlist that consists of very few tracks (cold-start problem), and (3) they neglect the context of a playlist such as the sequence of tracks or playlist title (context-aware continuation). This year's ACM RecSys Challenge'18 aimed to find new solutions to tackle these problems. In this paper, we propose a multimodal collaborative filtering model to deal effectively with diverse data. This consists of two components: (1) an autoencoder using both the playlist and its categorical contents and (2) a character-level convolutional neural network using the playlist title only. By simultaneously analyzing the playlist and the categorical contents, our model successfully addresses the cold-start and popularity bias problems. In addition, we consider the context of a playlist by utilizing its title, thus enhancing the prediction of well-suited tracks. In the challenge, our team "hello world!" was ranked the 2nd place, scoring 0.224, 0.394, and 1.928 for the three evaluation metrics, respectively. Our implementation code is publicly available at https://github.com/hojinYang/spotify_recSys_challenge_2018.