Komparasi Linear Regression, Random Forest Regression, dan Multilayer Perceptron Regression untuk Prediksi Tren Musik TikTok

Abstract

Prediksi bagaimana korelasi feature audio terhadap lagu yang populer di TikTok merupakan hal yang perlu dilakukan dalam industri musik. Dengan bekal data yang memiliki beberapa feature audio maka dilakukan penelitian menggunakan metode Linear Regression, Random Forest Regression (RFR), dan Multilayer Perceptron Regression (MLP Regression) untuk membandingkan model yang dapat memprediksi popularitas secara efektif dan feature yang mempengaruhi popularitas lagu di TikTok, kemudian dilakukan juga Exploratory Data Analysis (EDA) untuk mendapatkan insight data. Hasil dari proses EDA yaitu popularitas lagu terbanyak berada di range 40 - 80, durasi lagu antara 2 - 3 menit, feature loudness berkorelasi positif dengan energy, demikian juga antara artist_pop dan track_pop. Set feature importance pada model LR dan RFR untuk feature target track_pop adalah artist_pop, loudness, dan duration_ms. Metode LR memiliki hasil paling efektif diantara RFR dan MLP Regression untuk dataset yang dipakai, yaitu dengan hasil MSE sebesar 0.0313, RMSE sebesar 0.177, dan MAE sebesar 0.118.

Predicting how audio features correlate with songs that are popular on TikTok is something that needs to be done in the music industry. Armed with data that has several audio features, a study was conducted using the Linear Regression, Random Forest Regression (RFR), and Multilayer Perceptron Regression (MLP Regression) methods to compare models that can effectively predict popularity and features that influence song popularity on TikTok, then Exploratory Data Analysis (EDA) was also carried out to gain insight into the data. The results of the EDA process are that the most popularity of songs is in the range of 40 - 80, the duration of songs is between 2 - 3 minutes, feature loudness is positively correlated with energy, so is between artist_pop and track_pop. The set feature importance in the LR and RFR models for the feature target track_pop is artist_pop, loudness, and duration_ms. The LR method has the most effective results between RFR and MLP Regression for the dataset used, namely with the results of MSE of 0.0313, RMSE of 0.177, and MAE of 0.118.