Using Deep Autoencoders to Improve the Accuracy of Automatic Playlist Generation

Manual creation of playlists is a trivial and time consuming task due to huge catalogue of songs available online. Automatic Playlist Generation systems are now available on various music platforms such as Spotify, Youtube, etc. Personalized playlists are generated by these systems considering likes and dislikes of users. These systems are developed by following two main approaches: Collaborative Approach and Content-based Approach. Systems based on collaborative approach requires enough user's data to generate accurate results while the accuracy of playlists generated using Content-based approach systems depend highly on features used from the dataset as well as learning algorithm. In this paper, the authors propose a hybrid model which leverages the benefits of above mentioned approaches. The proposed system uses Deep Autoencoders to extract the most important features of songs present in the dataset to form clusters. Data about user's playlist i.e. User-level data is combined with the clusters to build the hybrid model. This approach is validated with the help of prototype implementation and a survey of 100 users.