Temporal Dynamics of the Last.fm Music Platform

Last.fm is a social networking platform established in 2002. According to Last.fm, the platform has over 20 million users on the site every month, which are based in more than 200 countries. After a user signs up, Last.fm records - among others - all artists a user listens to, aggregates this information over seven days and provides lists of the most listened artists for each week over the lifetime of a user. We use this information to build a user's profile by extracting the genres of the most listened artists. The artist's genre is determined by the tags that the community members use to characterize the artist. We represent each users as vertices in a graph and connect users with an edge, if their profile similarity reaches a predefined threshold. The similarity is determined by calculating the distance between pairs of genre vectors using the cosine similarity measure.

We randomly chose approx. 600,000 users and obtained over a period of 167 weeks (September 2005 to November 2008) their weekly artists charts. Since many users are not active on a regular basis, we chose randomly 2,000 users from this set who were active in at least 80% of all periods.

We applied DenGraph-IO on the resulting graph to detect and observe the evolution of clusters during the observation period of 115 weeks. The aim was to see, whether the proposed clustering techniques detects meaningful communities and evolutions[1][2]. In the following we focus on 5 weeks to analyze the music listening behavior. There temporal dynamic and the cluster transitions are shown at the right side.

All five clusterings consist of a giant component and six to nine clusters of different sizes. Therefore, the average cluster size is very high (126 members). The large component can be explained by the user structure of the data set: We observe a large amount of very similar user profiles. These users have a profile vector with a large weight in the genre indie, indie rock and alternative. This cluster is very strong over all intervals and the number of members varies only slightly over all periods. Besides this cluster, smaller clusters consisting of users with a different listening behavior are detected. The graph visualizations and a more detailed description of the five clusterings are shown below.

Last.fm Interval 106

In week 8/2007, nine clusters are detected. The biggest cluster is the indie group. Furthermore, the clusters which represent the six main genres are displayed: the indie, the hip-hop, the metal, the industrial, the j-pop/j-rock and the rock cluster. Besides, a small cluster is detected which is labeled with tags that are also used for labels of three other clusters (punk, metal, rock). With one of these cluster (indie) exists an overlap and edges exit also to the other two clusters (rock, pop and metal). This indicates a closeness of the cluster to three other clusters and, in fact, in the next interval a merge of this cluster with one of the other clusters can be observed.

In week 9/2007, a merge of the punk, metal, rock cluster with the larger metal cluster occurs. One small cluster (pop, rnb, soul) has been removed. Some members of the rock, pop cluster split into a rock cluster which overlaps with the indie cluster. The other five clusters (indie, industrial, hip-hop, j-pop, and electronic) are unchanged.

Last.fm Interval 108

From week 9/2007 to 10/2007 three cluster removals: Two of the removed clusters will show up again in the next interval (electronic and rock) and the cluster which is labeled with rock and pop is not detected in later periods. Furthermore, a split of the j-pop, j-rock cluster is observable. It splits into actually three clusters: one j-pop cluster and two overlapping j-rock clusters (which will be merged in the next interval).

Last.fm Interval 111

In week 13/2007 ten clusters are detected. Eight of them have been discovered in earlier periods and two are new. Both two new clusters result from a split. The industrial, electronic cluster splits into an industrial, rock and an industrial, pop cluster. The second split concerns the metal cluster where a smaller metal, rock cluster is separated from the larger metal cluster.

Last.fm Interval 112

From week 13/2007 to 14/2007 one split, one merge and a removal take place. The two industrial clusters are merged to one and the electronic cluster is removed. From the large metal cluster two clusters are separated. One cluster is labeled metal, heavy metal, rock and the second cluster is labeled with the tags punk, metal, a cluster already observed in week 8/2007 which merged in week 9/2007 with the larger metal cluster. Furthermore, the cluster metal, rock, alternative has no longer an overlap with the metal cluster.


  1. Schlitter N, Falkowski T. Mining the Dynamics of Music Preferences from a Social Networking Site. In: Proceedings of the 2009 International Conference on Advances in Social Network Analysis and Mining. Athens: IEEE Computer Society; 2009. p. 243-8.
  2. Falkowski T, Schlitter N. Analyzing the Music Listening Behavior and its Temporal Dynamics Using Data from a Social Networking Site. Zurich; 2008.