Inverted index techniques in music information retrieval: three application scenarios

• Ano de defesa: 2023
• Autor(a) principal: Tofani, Arthur Piza Mosterio
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Audio identification; Content-based retrieval; Identificação de áudio; Identificação de versão; Indexação; Indexing; Information retrieval; Music information retrieval; Ordenação por relevância; Recommender systems; Recuperação de informação; Recuperação de informação musical; Scoring techniques; Sistemas de recomendação; Version identification
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-07102024-212635/

The rapid advancement of digital technology has resulted in the widespread use of search engines and recommender systems across several industries. In music streaming platforms, these systems are transforming the way we discover and experience music. However, the growing number of users and items available in catalogues represents a significant challenge for these platforms. The Information Retrieval (IR) research field has a long-standing history of scale-related issues, initially driven by text search needs. Over time, this area has evolved to address a wide range of retrieval requirements, offering a range of techniques that can be employed for solving problems in many contexts. Indexing and searching are two complementary activities that form the foundation of IR systems. These systems are usually supported by the inverted index, a central data structure that enables the efficient indexing and retrieval of relevant information. Designing IR systems for non-textual scenarios is not a trivial task. For instance, the accurate scoring and ranking of the retrieved results requires representing the problem as a collection of indexable units. In the field of Music Information Retrieval (MIR), numerous tasks rely on non-textual representations (such as digital audio and music data), presenting distinct challenges for efficient retrieval. This study explores the application of IR techniques for music retrieval through the investigation and assessment of three distinct music tasks where scaling is an important requirement: audio identification, version identification, and music recommendation. The study aims to develop a better understanding of the challenges involved in designing effective IR systems for music purposes, with a particular focus on understanding the effectiveness of typical IR scoring techniques in these tasks. The findings of this study suggest that typical text-based scoring techniques behave as competitive baselines for relevance-based ranking, enabling researchers to focus on representation aspects. Also, the study highlights the importance of evaluating the number of results returned by retrieval systems along with the performance. Also, the study highlights the impact of the number of retrieved items in scalability and efficiency.