Personalised music playlist continuation

Abstract

With the rise of digital music consumption, there is a growing demand for personalised and specific playlists, targeting particular feelings or situations. However, state-of-the-art music playlist generation and continuation models face difficulties in adapting to specific user preferences. While they manage to successfully recommend songs to the user based on their existing playlist, they do not allow the user the liberty of specifying whether they want the recommended songs to be popular, recently released, mostly instrumental, or otherwise. This research presents a new potential music playlist continuation framework, which can enhance music playlist continuations by defining importance weights for each personalisation. In particular, we allow the user to specify importance weights for song popularity, acousticness, danceability and instrumentalness. In our research, we transform our dataset into a knowledge graph, on which knowledge graph embeddings are applied. By doing so, we transform every component of our knowledge graph into a vectorised feature, allowing us to perform link prediction. By performing link prediction, the solution developed can recommend potentially suitable songs to be added to a given playlist. If our personalisation aspect is applied, the recommended songs will be ranked according to how much they satisfy the importance weights provided by the user. Through our work, we find which is the most suitable knowledge graph embedding for our solution, by comparing the results of different experiments which utilise the TransE, RESCAL and DistMult knowledge graph embeddings. We find that for our dataset and solution, our TransE models seems to be the most effective. Furthermore, we perform feature ablation studies, in order to compare the results obtained when certain features from our data are excluded. We produce an innovative framework which helps us expand existing research targeting music playlist continuation and personalisation. Furthermore, we attempt to apply weighting to knowledge graph triples, which at the time of writing, is still an open research question.