Mūzikas klasifikācijas algoritmu nepilnības, izmantojot ģenerētus lietošanas gadījumus

Abstract

Bakalaura darbā "Muzikālo īpašību kopas labākai žanru klasifikācijas precizitātei" tiek apskatīti mūzikas klasifikācijas algoritmi un izvēlētas muzikālās īpašības, ar kurām klasifikācija tiek veikta. Tiek veikta mūzikas īpašību selekcija, lai uzlabotu klasifikācijas algoritmu precizitāti, izmantojot divas klasifikācijas metodes – K-Nearest neighbors un Support vector machines. Izmantojot pilnu kopu ar muzikālajām īpašībām, tiek sasniegta visaugstākā precizitāte – ar selekcijas metodēm tā tikai pasliktinās. Tiek uzsvērts, ka turpmākajos pētījumos būtu jāiekļauj plašāka īpašību un metožu izvēle, lai uzlabotu algoritmu efektivitāti. Diplomdarbs ir uzrakstīts angļu valodā, tā apjoms ir 37 lapaspuses, tajā ir iekļautas 13 figūras, 1 pielikums, un izmantoti 26 literatūras avoti. Atslēgas vārdi: mūzikas klasifikācija, īpašību kopas, algoritmi, selekcija, precizitāte.

Executive summaryAs the music industry is rapidly expanding, so are the needs of consumers. Many modern music streaming platforms use music classification algorithms to provide categorized, personalized content to end-users. This research paper assesses understanding through musical features, which are numerical values. Therefore, musical feature extraction is performed, where music is measured in several defined variables or feature vectors. A combination or set of feature vectors contributes to a “fingerprint” of a particular musical sample. This data is used to train and test the algorithms for classifying music by genre. This research uses two algorithms: K-Nearest Neighbors and Support Vector Machines. Both algorithms are first trained and tested by using all chosen musical features. Then, two existing methods are taken to seek a subset of features that would outperform the entire set. Lastly, a third method aims to find a subset of musical features that would increase accuracy in one or both algorithms. When the entire set of musical features was applied to the models, a reasonably high accuracy was obtained for both models (over 80%). However, when using the three feature selection methods, the accuracy only decreased. Moreover, experimenting with the three feature selection methods found that an increase in musical feature vectors leads to increased accuracy of the models. One of the limitations is that only three feature selection methods were used, and only data from one dataset was used. A more comprehensive selection of features, more feature selection methods, and more algorithms should be included in the scope of further research.