A Graph-Based Method for Playlist Generation Debora C. Correa 1 , Alexandre L. M. Levada 2 and Luciano da F. Costa 1 1 Instituto de Fisica de Sao Carlos, Universidade de Sao Paulo, Sao Carlos, SP, Brazil 2 Departamento de Computacao, Universidade Federal de Sao Carlos, SP, Brazil deboracorrea@ursa.ifsc.usp.br, alexandre@dc.ufscar.br, luciano@ifsc.usp.br Abstract. The advance of online music libraries has increased the im- portance of recommendation systems. The task of automatic playlist generation naturally arises as an interesting approach to this problem. Most of existing applications use some similarity criterion between the songs or are based on manual user interaction. In this work, we pro- pose a novel algorithm for automatic playlist generation based on paths in Minimum Spanning Trees (MST’s) of music networks. A motivation is to incorporate the relationship between music genres and expression of emotions by capturing the presence of temporal rhythmic patterns. One of the major advantages of the proposed method is the use of edge weights in the searching process (maximizing the similarity between sub- sequent songs), while Breadth-First (BF) and Depth-First (DF) search algorithms assume the hypothesis that all the songs are equidistant. Keywords: playlist, rhythm, graphs, search algorithms. 1 Introduction With the dissemination of online resources with music content, music recommen- dation systems have received much attention. Indeed, the sometimes manual and time-consuming selection task of music playlists can be replaced by automatic al- gorithms. Such algorithms can generate the playlist according to the user’s music preferences or through some defined similarity criterion between the songs. We can relate three main important aspects of a playlist: the individual songs themselves, the order in which they are played, and the size of the playlist. In the literature, we can find earlier efforts concerning the automatic generation of musical playlists [1, 10, 7, 9]. Most of these approaches are based on collabora- tive filtering techniques, audio content analysis, and require a manually labeled database or the analysis of metadata. In this scenario, our contribution is to propose a novel and unsupervised playlist algorithm, avoiding possible noise due to the user’s subjectiveness. Be- sides, a motivation is the possibility to associate music genres to the presence of temporal patterns in the rhythm as a way to express notions of emotion. According to [12], regular and smooth rhythmic patterns indicate expressions like happiness, joyness or peacefulness. Irregular and complex rhythms indicate 9th International Symposium on Computer Music Modelling and Retrieval (CMMR 2012) 19-22 June 2012, Queen Mary University of London All rights remain with the authors. 466