Back to search

Institutional Repository of Vilnius University

Muzikos rekomendacijų informacinė sistema /

Abstract

dc:description

With the increasing volume of digital music catalogs, it is increasingly difficult for users to independently find music content that is relevant to them. For this reason, music recommendation systems become a necessary tool for automated evaluation of the similarity of music tracks with each other. This thesis examined how vector representations of music tracks are formed through different methods of processing and extracting sound signals and how the quality of music recommendations depends on the selected models, signal representations and chosen data sets. which is important in a music recommendation information system. The purpose of this thesis was to create a music recommendation system — an app, which, based on the analysis of the audio signal, would generate a list of recommendations for music tracks similar to the loaded audio track for the user. To achieve this, the tasks of literature analysis, selection of datasets, calculations of algorithms and building a system, were carried out. One created and specifically trained MRN model and two pretrained models, OpenL3 and CLAP, were used. In order to evaluate the performance of models with different datasets, GTZAN and GTZAN joined together with FMA Small datasets were used to extract model audio embeddings. The comparative experiment consisted of several parts involving the models. Part of the MusicRecNet model included model creation, training, processing of raw data and transformation into mel—spectrograms, converting signal representation from mel—spectrograms to embeddings. The preparation and use of OpenL3 and CLAP models differed from converting MusicRecNet’s audio signal into embeddings, as these models did not need to convert the data to another representation — that process was carried out in the internal architecture of the models. Each model was used to extract embeddings from the GTZAN and GTZAN + FMA small datasets. After that, a comparison between the embeddings was conducted using a KNN algorithm and the calculation of the metrics, cosine, euclidean distances. By comparing the models and different datasets recommendations results were concluded. After completing the tasks, results show that the MRN model with both of the datasets can perform the best embedding extraction and gives the most accurate recommendation of similar audio tracks.

Degree

thesis:*
Grantor dc:publisher
Institutional Repository of Vilnius University
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Matiukaitė, Viktorija,

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
lit

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Matiukaitė, Viktorija,. Muzikos rekomendacijų informacinė sistema /. Institutional Repository of Vilnius University, 2026. https://repository.vu.lt/VU:ELABAETD277689974&prefLang=en_US