| dc.contributor.advisor | Erfiani | |
| dc.contributor.advisor | Rahardiantoro, Septian | |
| dc.contributor.author | Ramadhan, Hafidz Candra | |
| dc.date.accessioned | 2026-08-05T01:54:43Z | |
| dc.date.available | 2026-08-05T01:54:43Z | |
| dc.date.issued | 2026 | |
| dc.identifier.uri | http://repository.ipb.ac.id/handle/123456789/177186 | |
| dc.description.abstract | Media sosial menjadi ruang utama bagi publik untuk mengekspresikan opini
terhadap kebijakan pemerintah, termasuk program Magang Nasional Kementerian
Ketenagakerjaan. Penelitian ini bertujuan mengevaluasi pelabelan otomatis GPT
4o mini dengan strategi zero-shot dan few-shot terhadap pelabelan manual,
membandingkan performa peramalan Long Short-Term Memory (LSTM)
berdasarkan sumber pelabelan terbaik dengan dua representasi input berupa
persentase sentimen positif harian dan persentase sentimen positif harian yang
dihaluskan dengan moving average 7 hari (MA-7), serta memproyeksikan sentimen
positif publik terhadap program tersebut. Sebanyak 7.240 tweet berbahasa
Indonesia dikumpulkan dan dilabeli secara manual oleh dua annotator dengan
Cohen's Kappa sebesar 0,7545. Strategi few-shot mengungguli zero-shot dengan
balanced accuracy 0,9012, Cohen's Kappa 0,8818, dan macro F1score 0,9199. Pada
tahap peramalan, input MA-7 secara konsisten mengungguli input harian mentah
dengan penurunan RMSE uji sebesar 81,1% pada pelabelan manual dan 61,2% pada
pelabelan few-shot. Proyeksi 30 hari ke depan menunjukkan kecenderungan
melemah pada akhir periode. Temuan ini mengindikasikan bahwa pelabelan few
shot dapat menjadi alternatif pelabelan manual yang efisien dan smoothing MA-7
bisa meningkatkan kinerja peramalan pada data sentimen harian yang berfluktuasi
tinggi. | |
| dc.description.abstract | Social media has become the primary space for the public to express
opinions on government policies, including the Ministry of Manpower's National
Internship program. This study aims to evaluate the automatic labeling of GPT-4o
mini using zero-shot and few-shot strategies against manual labeling, compare the
forecasting performance of Long Short-Term Memory (LSTM) based on the best
labeling source with two input representations namely the daily positive sentiment
percentage and the daily positive sentiment percentage smoothed with a 7-day
moving average (MA-7), and project public positive sentiment toward the program.
A total of 7,240 Indonesian-language tweets were collected and manually labeled
by two annotators, achieving a Cohen's Kappa of 0.7545. The few-shot strategy
outperformed the zero-shot strategy, yielding a balanced accuracy of 0.9012, a
Cohen's Kappa of 0.8818, and a macro F1score of 0.9199. In the forecasting stage,
the MA-7 input consistently outperformed the raw daily input, showing a decrease
in test RMSE by 81.1% for manual labeling and 61.2% for few-shot labeling. A 30
day ahead projection indicates a weakening trend toward the end of the period.
These findings indicate that few-shot labeling can serve as an efficient alternative
to manual labeling, and MA-7 smoothing can enhance forecasting performance on
highly fluctuating daily sentiment data. | |
| dc.description.sponsorship | | |
| dc.language.iso | id | |
| dc.publisher | IPB University | id |
| dc.title | Peramalan Sentimen Publik terhadap Magang Nasional pada Media Sosial X Menggunakan LSTM dengan Pelabelan Large Language Model | id |
| dc.title.alternative | | |
| dc.type | Skripsi | |
| dc.subject.keyword | Large Language Model | id |
| dc.subject.keyword | LSTM | id |
| dc.subject.keyword | Moving Avarage | id |
| dc.subject.keyword | Peramalan Sentimen | id |
| dc.subject.keyword | Pelabelan Otomatis | id |
| dc.subtype | Undergraduate Theses | |