Analisis Sentimen Pengguna X terhadap Pelaksanaan Program Mbg menggunakan Metode Multinomial Naive Bayes

Abstract

Program Makan Bergizi Gratis (MBG) merupakan inisiatif pemerintah untuk menyediakan makanan bergizi secara gratis kepada pelajar, balita, ibu hamil, dan menyusui, namun efektivitasnya memerlukan evaluasi respons masyarakat. Penelitian ini bertujuan menggunakan algoritma Multinomial Naive Bayes untuk menganalisis sentimen komentar pengguna platform X terhadap pelaksanaan MBG. Algoritma MNB digunakan karena kemampuannya dalam menangani matriks fitur sparse serta mengolah data bervolume besar secara cepat, akurat, dan transparan. Data dikumpulkan pada periode awal pelaksanaan program (periode baseline) dan setelah program berjalan beberapa bulan (periode endline). Sebanyak 6875 komentar yang memuat istilah “MBG” dikumpulkan pada bulan Januari 2025 (periode baseline) dan bulan April 2025 (periode endline). Komentar tersebut diberikan label positif, netral, atau negatif melalui hybrid AI-human annotation, serta diproses meliputi pembersihan teks, tokenisasi, normalisasi, stopword removal, dan stemming. Fitur diekstraksi dengan menggunakan metode term frequency (TF). Model MNB dengan back-translation menghasilkan weighted F1-score tertinggi sebesar 0,79. Model tersebut menunjukkan kinerja unggul dalam mendeteksi sentimen positif (F1 = 0,84) dan negatif (F1 = 0,83), walaupun klasifikasi netral masih terbatas (F1 = 0,58). Analisis temporal mengungkap pergeseran dominasi sentimen negatif (55 %) menjadi positif (58 %) pada tiga bulan setelah implementasi program. Temuan ini menegaskan efektivitas pendekatan MNB untuk pemantauan sentimen skala besar.

The Free Nutritious Meal Program (MBG) is a government initiative to provide free nutritious meals to students, toddlers, pregnant, and lactating women, yet its effectiveness requires evaluation of public response. This study aims to use the Multinomial Naive Bayes algorithm to analyze the sentiment of X user comments on MBG implementation. The MNB algorithm is used for its ability to handle sparse feature matrices and process large volumes of data quickly, accurately, and transparently. Data are collected during the initial period of program implementation (baseline period) and after the program has been running for several months (endline period). A total of 6,875 comments containing the term MBG were collected in Januari 2025 (baseline period) and April 2025 (endline period). Comments were given labels as positive, neutral, or negative through hybrid AI-human annotation and processed via text cleaning, tokenization, normalization, stopword removal, and stemming. Features were extracted using term frequency (TF). The back-translation MNB model achieved a highest weighted F1-score of 0,79. The model demonstrate strong performance in detecting positive (F1 = 0,84) and negative sentiments (F1 = 0,83), although neutral classification remained limited (F1 = 0,58). Temporal analysis revealed a shift from predominantly negative sentiment (55%) to predominantly positive sentiment (58%) after three months of program implementation. These findings confirm the effectiveness of the MNB approach for large-scale sentiment monitoring.