Optimizing Recommendation Algorithms Using Self-Similarity Matrices for Music Streaming Services

Most music recommendation algorithms such as proprietary ones designed for curating music in music streaming companies rely mainly on song ratings data to be able to recommend songs to users. This does not always provide a great music experience for many users who pay for these services. As an example there are over 35 million songs on Spotify alone and according to recent statistical report released by Spotify there are about 4 million songs on Spotify that have never been played not even once. Clearly this is not good news to content creators who trust and pay Spotify as a platform to distribute their music. The root of the problem lies on the music recommendation algorithm adopted and used by many music streaming platforms like Spotify. The algorithm fails to recommend great music to the users which they can enjoy and this results in the platform experiencing “dark music” or music that has never received any play in the platform. We propose a framework which provides a unique strategy that can be adopted by music recommendation algorithms to give users a better music experience. We utilize self-similarity matrices developed in R to visualize patterns of repetition in text extracted song lyrics. The way this works is that the lyrics of a song played by the user are extracted and a visual pattern print of a song is generated and this pattern is then compared with many other patterns of songs existing on the platforms especially ones that have not been played before. A comparison is then performed and if the comparison similarity index is above 70 percent then a song is recommended to the user. A song with high a similarity index gets first priority. We believe this will ensure great music experiences for the listeners and also benefit content creators since the likelyhood that their music will reach users will be high.