features and evaluate our method on other databases,
because we do not know what type of manifold
structure under the data sets would be more sensitive
to the ranking algorithm. Moreover, many other RF
approaches should be considered. Because MR has its
own drawbacks in handling large-scale data sets[25],
another future extension of our work will be to improve
its diversity at high ranks. In this study, genre is the
only musical topic considered and used to judge tracks
as relevant or irrelevant. However, in actuality, audio
content has multiple topics and many works[25–27] have
explored the problem of diversity in document and
image retrieval; we also should consider the problem
of high-ranking diversity in large-scale music retrieval.