A Combined Music Label Propagation Model
Jing Cai, Heng Li, Bo Lang · 2011
Music labels, especially those related to high level semantics are very useful in music retrieval and recommendation, but normally hard to acquire. In the submission to ISMIR'07, Mohamed Sordo proposed a novel model, i.e., propagation of labels, to annotate music with existing labels, by using the content-based music similarity distance. In that model, a partially annotated collection with a lot of non-labeled music was annotated at a high precision and recall. In this paper, we proposed a new model -- label probability prediction model -- and introduce it into the Sordo's work, which makes a combined model, to improve the accuracy of propagation without exploiting any other information. In addition, we also made some modifications to the original Sordo's model that could make the algorithm works better. Then we compare the result of combined model to that yielded by the original on a publicly accessible ground truth data, and find that, the new approach can reach a higher recall. Furthermore, with the same recall, our method obtains a better precision.