Skip to main content – Music Recommendation incorporating social network ties and collaborative filtering is a music recommendation and social networking site incorporating several different types of data. Each user can ‘scrobble’ songs they listen to on iTunes, Spotify, and various other web services and software. keeps track of which tracks have been scrobbled to each user’s profile, and finds patterns in users listening habits. Users can become friends with people they already, or people they get to know on the site. Based on which artists a user listens to the most, can also find musical ‘neighbors’ who listen to similar music, and use them to recommend new music a user might like.’s data network can be thought of as a graph with two basic types of edges: those from users to other users, and those from users to musical artists they are fans of. Links from user to user can either be actual friends of the user, or neighbors that has recommended based on similarity of musical taste. Links between user and artist can be stronger or weaker, depending on their listening frequency.’s recommendation engine, utilizing collaborative filtering and social influence data, can then attempt to predict artists a user might like but have not yet listened to, by closing open triads in the user’s local network. Collaborative filtering is a long used technique in recommender systems, and works by recommending based on other users with similar tastes (in this case, musical neighbors). has the additional advantage of having access to social network data, linking users who might not have as similar tastes (although this is quite possible) but are socially connected. The combination of these two types of data results in a very powerful recommendation engine.

Consider the following example. If user A is a neighbor of user B, and they have similar musical tastes, then there is a very strong tie between them. If user B is a big fan of artist C, and has scrobbled them numerous times, then there is also a strong tie between them. Based on’s data, user A has not yet listened to artist C (no link has formed between them yet), and there is a good chance that user A will also like artist C. If the recommendation is succesful, the gap in a strong triad will be filled.’s recommendation engine thereby utilizes the triadic closure property to recommend new music to it’s users. Recent research has shown that predicting based on voluntary friendships may be even more effective (Ye, Liu & Lee). While musical taste similarity can be a good predictor of future musical preferences, social influences can have an ven stronger influence on taste formation.

There are a number of factor that are necessary to make such a recommendation engine work. First and foremost, some starting data on musical preferences are needed. This is referred to as the cold start problem, as it can be very hard for recommender systems like these to work with new users who have little to no musical preference data in their profile. This problem is lessened by encouraging users to enter their favorite artists upon making a profile, connecting to facebook profiles and using their music data, and making it easy to start scrobbling their digital music listening right away. This system also needs a very large network of fans and artists, incorporating a diverse pool of musical tastes and listening habits to accomodate a large variety of users. Considering the huge diversity of music available on the internet today, a great deal of data is needed to accurately predict music for new users. Compared to the artist network as a whole, each user’s taste network is quite small, and it is difficult to build up taste data for all the subsets of the musical network.




Leave a Reply

Blogging Calendar

September 2012