An effective re-ranking method based on learning to rank for improving audio fingerprinting

This paper presents an effective re-ranking method that uses learning-to-rank paradigms to improve the accuracy of landmark-based audio fingerprinting (AFP) for audio music retrieval. The re-ranking mechanism is invoked whenever the returned ranking from an AFP system does not have a high enough confidence measure. We propose that use of new features for re-ranking, and employ the popular learning-to-rank paradigms, including pairwise and listwise approaches for modeling the behavior from queries to desired ranking. Experimental results indicate that the proposed re-ranking method can effectively improve the top-1 recognition rate of our AFP system, with only small extra overhead of overall response time.