ENHANCING MUSIC RECOMMENDATION PIPELINE THROUGH DEAM-BASED TRACK MOOD CLASSIFICATION FOR MOOD-ALIGNED AND COHERENT PLAYLIST GENERATION

Kemal Susanto, Dr. techn. Khabib Mustofa, S.Si.,M.Kom. ; Muhammad Oriza Nurfajri S.Kom.,M.IT.

2026 | Skripsi | ILMU KOMPUTER

Sistem rekomendasi musik berbasis content-based filtering umumnya merekomendasikan lagu berdasarkan kemiripan fitur audio, tanpa mempertimbangkan mood lagu atau menjaga kesinambungan musikal, seperti keselarasan nada dasar (key) dan tempo antar lagu. Akibatnya, playlist yang dihasilkan cenderung terasa generik, sementara transisi antar lagu menjadi kurang koheren. Penelitian ini bertujuan untuk meningkatkan alur rekomendasi musik dengan menggabungkan klasifikasi mood berbasis DEAM serta analisis key dan tempo untuk menghasilkan playlist yang mempertahankan relevansi terhadap target mood dan koherensi transisi musikal. Metode yang digunakan dalam penelitian ini meliputi ekstraksi fitur MFCC, pembentukan label mood berdasarkan model valence-arousal, pelatihan model CNN-LSTM, serta pengurutan lagu menggunakan key compatibility dan tempo smoothness. Model klasifikasi memperoleh akurasi rata-rata sebesar 0.5756, macro precision sebesar 0.4494, macro recall sebesar 0.4546, dan macro F1-score sebesar 0.4442. Pada evaluasi playlist yang dilakukan selama 9 run dan kemudian dirata-ratakan, metode yang diusulkan menghasilkan nilai Precision@10 sebesar 0.0556 untuk Anxious, 0.2778 untuk Calm, 0.5000 untuk Happy, dan 0.8333 untuk Sad. Metode yang diusulkan juga menghasilkan Mean Circle-of-Fifths Distance sebesar 0.2593–0.6543, Mean Absolute Tempo Difference sebesar 4.5854–13.3507 BPM, dan Standard Deviation of Tempo Difference sebesar 3.5608–6.9831 BPM. Hasil ini menunjukkan bahwa metode yang diusulkan mampu memberikan kompromi yang seimbang antara relevansi terhadap target mood dan koherensi transisi musikal.

Music recommendation systems based on content-based filtering commonly recommend songs according to audio feature similarity, without considering the song’s mood or maintaining musical continuity, such as key and tempo alignment between songs. As a result, the generated playlists tend to feel generic, while song transitions may become less coherent. This research aims to enhance the music recommendation pipeline by combining DEAM-based mood classification with key and tempo analysis to generate playlists that maintain target-mood relevance and musical transition coherence. The method used in this research includes MFCC feature extraction, mood label construction based on the valence-arousal model, CNN-LSTM model training, and song ordering using key compatibility and tempo smoothness. The classification model achieved an average accuracy of 0.5756, a macro precision of 0.4494, a macro recall of 0.4546, and a macro F1-score of 0.4442. In the playlist evaluation, which was conducted over 9 runs and then averaged, the proposed method achieved Precision@10 values of 0.0556 for Anxious, 0.2778 for Calm, 0.5000 for Happy, and 0.8333 for Sad. The proposed method also achieved Mean Circle-of-Fifths Distance values ranging from 0.2593 to 0.6543, Mean Absolute Tempo Difference values ranging from 4.5854 to 13.3507 BPM, and Standard Deviation of Tempo Difference values ranging from 3.5608 to 6.9831 BPM. These results indicate that the proposed method provides a balanced compromise between target-mood relevance and musical transition coherence.

Kata Kunci : Music recommendation, mood classification, DEAM, CNN-LSTM, playlist coherence, content-based filtering.