Tag-Aware Spectral Clustering of Music Items.

Socialtaggingis anincreasingly popularphenomenon with substantial impact on Music Information Retrieval (MIR). Tags express the personal perspectives of the user on the music items (such as songs, artists, or albums) they tagged. These personal perspectives should be taken into account inMIRtasksthatassessthesimilaritybetweenmusicitems. In this paper, we propose an novel approach for cluster- ing music items represented in social tagging systems. Its characteristic is that it determines similarity between items by preserving the 3-way relationships among the inherent dimensions of the data, i.e., users, items, and tags. Con- versely to existing approaches that use reductions to 2- way relationships (between items-users or items-tags), this characteristic allows the proposed algorithm to consider the personal perspectives of tags and to improve the clus- tering quality. Due to the complexity of social tagging data, we focus on spectral clustering that has been proven effec- tive in addressing complex data. However, existing spectral clustering algorithms work with 2-way relationships. To overcome this problem, we develop a novel data-modeling scheme and a tag-aware spectral clustering procedure that uses tensors (high-dimensional arrays) to store the multi- graph structures that capture the personalised aspects of similarity. Experimental results with data from Last.fm in- dicate the superiority of the proposed method in terms of clustering quality over conventional spectral clustering ap- proaches that consider only 2-way relationships.