A Probabilistic Model to Combine Tags and Acoustic Similarity for Music Retrieval

The rise of the Internet has led the music industry to a transition from physical media to online products and services. As a consequence, current online music collections store millions of songs and are constantly being enriched with new content. This has created a need for music technologies that allow users to interact with these extensive collections efficiently and effectively. Music search and discovery may be carried out using tags, matching user interests and exploiting content-based acoustic similarity. One major issue in music information retrieval is how to combine such noisy and heterogeneous information sources in order to improve retrieval effectiveness. With this aim in mind, the article explores a novel music retrieval framework based on combining tags and acoustic similarity through a probabilistic graph-based representation of a collection of songs. The retrieval function highlights the path across the graph that most likely observes a user query and is used to improve state-of-the-art music search and discovery engines by delivering more relevant ranking lists. Indeed, by means of an empirical evaluation, we show how the proposed approach leads to better performances than retrieval strategies which rank songs according to individual information sources alone or which use a combination of them.