MSALSA - a method of positioning search results in music information retrieval systems

Although the music information retrieval is already well described topic it still remains as a research subject of many scientists. Recently popular MP3 search engines are currently replaced with more advanced music information retrieval systems, that provide new ways of creating queries by user – like query by humming or query by singing. The files returned as a result system response are ordered with selected ranking algorithm. Because audio files are different data source than text, it give a new way to calculate final rank of music file, returned as a result. In fact there are query dependent (dynamic) and query independent (static) measures. Query dependent measures are based on relevance between information given by user as a query input data and the meta data stored in a file. Meta data can also be used in calculation of static measures, like relevance between file meta data and its original values. Although meta data are very useful, it do not provide any information about music file context for example: if the audio file is referenced from the popular Website. That causes necessity of 540 Z. Mazur, K. Wiklak applying Web structure algorithms, that use connections (links) between sites as a primary knowledge source about Websites. Such algorithms are HITS, PageRank, SALSA. After short analysis of pros and cons of each, the PageRank have been chosen as a base of developing new algorithm, that uses also music file links – MusicPageRank. Section 1 is a short introduction. Section 2 describes result music files ranking problem, including query dependent and independent measures. This section provides a general formula of calculating static FileRank measure and describes role of meta data, contextual file information and individual preferences in calculating FileRank. Section 3 describes methods of modification link based algorithms to work with music information retrieval systems. The Web is described as a Directed graph of connection between Websites (Fig 1) and as a site Adjacency Matrix (1). A transformation from Adjacency Matrix to Reduced Adjacency Matrix (2), which is build using Websites, that contain links to music files is also described. Section 4 describes a proposed MusicPageRank algorithm (MPR). The MPR is discussed using example graph on (Fig. 2), which corresponds to Reduced Adjacency Matrix (3). The calculation of result rank vector is described with using a iterative power method. This section ends with MPR test results, presented on (Fig 3). The tests results have prooved that MPR is fast convergent algorithm, that can be applied in real music search engines.