PENERAPAN CONVOLUTIONAL NEURAL NETWORK DALAM PENGENALAN EMOSI BERBASIS EEG UNTUK SISTEM REKOMENDASI MUSIK
Main Article Content
Abstract
Emosi merupakan aspek penting dalam kehidupan manusia yang memengaruhi perilaku, pengambilan keputusan, dan kesejahteraan mental. Dalam era digital, kebutuhan akan sistem yang mampu mengenali emosi secara real-time semakin meningkat. Penelitian ini bertujuan untuk mengembangkan sistem rekomendasi musik berbasis emosi menggunakan sinyal EEG (Electroencephalogram) dan metode Convolutional Neural Network (CNN). Model CNN dirancang dengan dua lapisan konvolusional dan pooling untuk mengekstraksi fitur dari sinyal EEG, diikuti oleh lapisan dense untuk mengklasifikasikan tiga kondisi emosi yaitu relax, excited, dan stressed. Hasil pelatihan menunjukkan bahwa model mampu mengenali pola dari data EEG, namun mengalami penurunan akurasi validasi setelah beberapa epoch, dengan akurasi tertinggi sebesar 60% yang kemudian menurun dan stagnan pada kisaran 43%. Gejala overfitting ringan juga teridentifikasi. Meskipun teknik EarlyStopping dan ReduceLROnPlateau digunakan untuk mengurangi risiko overfitting, performa model masih perlu ditingkatkan agar dapat diterapkan secara efektif dalam sistem rekomendasi musik berbasis kondisi afektif pengguna.
Emotions are a crucial aspect of human life, influencing behavior, decision-making, and mental well-being. In the digital era, the demand for systems capable of recognizing emotions in real-time is increasing. This study aims to develop an emotion-based music recommendation system using EEG (Electroencephalogram) signals and the Convolutional Neural Network (CNN) method. The CNN model is designed with two convolutional and pooling layers to extract features from EEG signals, followed by dense layers to classify three emotional states: relaxed, excited, and stressed. The training results indicate that the model can recognize patterns in EEG data; however, a decline in validation accuracy was observed after several epochs, with the highest accuracy reaching 60% before dropping and stagnating around 43%. Signs of mild overfitting were also identified. Although techniques such as EarlyStopping and ReduceLROnPlateau were employed to reduce the risk of overfitting, the model's performance still requires improvement for effective implementation in a music recommendation system based on users' affective states.
Article Details
Section
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.