Identifikasi Kelompok Kata yang Berasosiasi pada Sentimen Masyarakat Terhadap Redenominasi Rupiah Menggunakan Generalized LASSO

Abstract

Kebijakan redenominasi rupiah kembali menjadi perhatian publik seiring diterbitkannya Peraturan Menteri Keuangan Nomor 70 Tahun 2025. Pemahaman terhadap sentimen masyarakat merupakan salah satu faktor yang dapat mendukung implementasi kebijakan tersebut. Penelitian ini bertujuan mengidentifikasi pola sentimen serta kata dan kelompok kata yang berasosiasi dengan sentimen pada komentar pengguna YouTube mengenai kebijakan redenominasi rupiah. Data diperoleh melalui web scraping pada empat kanal berita selama periode November hingga Desember 2025 dan menghasilkan 6.676 komentar setelah prapemrosesan. Skor polaritas dihitung menggunakan TextBlob, sedangkan pembobotan fitur dilakukan menggunakan Term Frequency–Inverse Document Frequency (TF–IDF). Pemodelan dilakukan menggunakan generalized Least Absolute Shrinkage and Selection Operator (LASSO) dengan matriks penalti identitas untuk identifikasi kata tunggal dan matriks penalti berbasis co-occurrence untuk identifikasi kelompok kata. Parameter regularisasi ditentukan menggunakan Approximate Leave-One-Out Cross Validation (ALOCV) dan Generalized Cross-Validation (GCV). Hasil analisis menunjukkan bahwa sentimen netral memiliki proporsi terbesar (41,36%), diikuti positif (38,95%) dan negatif (19,70%). Berdasarkan hasil evaluasi, matriks penalti dengan threshold 15 dipilih sebagai model terbaik. Koefisien positif berkaitan dengan stabilitas ekonomi dan dukungan terhadap kebijakan, sedangkan koefisien negatif berkaitan dengan kekhawatiran terhadap korupsi dan kenaikan harga.

The rupiah redenomination policy has regained public attention following the enactment of Minister of Finance Regulation No. 70 of 2025. Understanding public sentiment can support the implementation of this policy. This study aimed to identify sentiment patterns and words or groups of words associated with sentiment in YouTube user comments on the rupiah redenomination policy. Data were collected through web scraping from four news channels from November to December 2025, resulting in 6,676 comments after preprocessing. Sentiment polarity scores were calculated using TextBlob, while feature weighting was performed using Term Frequency–Inverse Document Frequency (TF–IDF). Modeling was conducted using generalized Least Absolute Shrinkage and Selection Operator (LASSO), with an identity penalty matrix for identifying individual words and a co-occurrence-based penalty matrix for identifying word groups. The regularization parameter was determined using Approximate Leave-One-Out Cross-Validation (ALOCV) and Generalized Cross-Validation (GCV). Neutral sentiment accounted for the largest proportion (41.36%), followed by positive (38.95%) and negative (19.70%) sentiments. The penalty matrix with a threshold of 15 was selected as the best model. Positive coefficients were associated with economic stability and policy support, whereas negative coefficients were associated with concerns about corruption and rising prices.