Lyric-Based Music Recommendation

Abstract

Traditional music recommendation systems rely on collaborative filtering to recommend songs or artists. This is computationally efficient and performs well method but is not effective when there is limited or no user input. For these cases, it may be useful to consider content-based recommendation. This paper considers a content-based recommendation system based on lyrical data. We compare a complex network of lyrical recommendations to an equivalent collaborative filtering network. We used user generated tag data from Last.fm to produce 23 subgraphs of each network based on tag categories representing musical genre, mood, and gender of vocalist. We analyzed these subgraphs to determine how recommendations within each network tend to stay within tag categories. Finally, we compared the lyrical recommendations to the collaborative filtering recommendations to determine how well lyrical recommendations perform. We see that the lyrical network is significantly more clustered within tag categories than the collaborative filtering network, particularly within small musical niches, and recommendations based on lyrics alone perform 12.6 times better than random recommendations.