Katamsi, Badar (2025) PERSONALISASI REKOMENDASI MUSIK UNTUK KEGIATAN BERKENDARA, BELAJAR, DAN MAKAN MENGGUNAKAN HYBRID MATRIX FACTORIZATION BERBASIS ALS DAN HDBSCAN PADA MILLION SONG DATASET. Undergraduate thesis, UPN Veteran Jawa Timur.
![[img]](https://repository.upnjatim.ac.id/style/images/fileicons/text.png) |
Text (Cover)
21081010186_COVER.pdf Download (1MB) |
|
Text (Bab 1)
21081010186_BAB I.pdf Download (461kB) |
|
Text (Bab 2)
21081010186_BAB II.pdf Restricted to Repository staff only until 15 September 2027. Download (835kB) |
|
Text (Bab 3)
21081010186_BAB III.pdf Restricted to Repository staff only until 15 September 2027. Download (1MB) |
|
Text (Bab 4)
21081010186_BAB IV.pdf Restricted to Repository staff only until 15 September 2027. Download (2MB) |
|
Text (Bab 5)
21081010186_BAB V.pdf Download (442kB) |
|
Text (Daftar Pustaka)
21081010186_DAFTAR PUSTAKA.pdf Download (343kB) |
|
Text (Lampiran)
21081010186_Lampiran.pdf Restricted to Repository staff only Download (257kB) |
Abstract
The development of music streaming platforms with catalogs of more than 100 millio n songs, such as Spotify, poses the challenge of information overload, making it diffic ult for users to find music that suits their preferences, while the cold start problem for new users also limits the accuracy of the recommendation system. This study aims to develop a contextual music recommendation system capable of addressing both issues by integrating Matrix Factorization based on Alternating Least Squares (ALS) and the HDBSCAN algorithm as clustering methods, along with audio feature-based filters to tailor recommendations to the user's activity context, namely driving, studying, and eating. The dataset used is the Million Song Dataset from Spotify & Last.fm, comprising 50,683 songs and nearly one million user listening histories. The research stages include data cleaning, transforming playcount into explicit ratings, extracting genre features using TF-IDF, dimension reduction with UMAP, clustering using HDBSCAN, and training the ALS model on each cluster. Evaluation results showed a Silhouette Score of 1.00, indicating optimal clustering, while the best ALS model achieved a Reconstructio n Error of 0.000814, RMSE of 0.1525, and MAE of 0.0630. A user satisfaction survey also revealed an average score of 4.2 out of 5, with 80% of recommended songs saved to playlists. Compared to the traditional single ALS-based approach, the proposed hybrid method proved more effective in addressing sparsity, cold start, and improving the relevance of occasion-based recommendations. In conclusion, the integration of ALS and HDBSCAN with audio feature filters can produce an efficient, accurate, and more personalized music recommendation system, thereby contributing to improved user experience quality on digital music platforms.
| Item Type: | Thesis (Undergraduate) | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
||||||||||||
| Subjects: | T Technology > T Technology (General) | ||||||||||||
| Divisions: | Faculty of Computer Science > Departemen of Informatics | ||||||||||||
| Depositing User: | Badar Katamsi | ||||||||||||
| Date Deposited: | 15 Sep 2025 06:05 | ||||||||||||
| Last Modified: | 15 Sep 2025 06:05 | ||||||||||||
| URI: | https://repository.upnjatim.ac.id/id/eprint/43527 |
Actions (login required)
 |
View Item |