Please use this identifier to cite or link to this item: http://repository.ipb.ac.id/handle/123456789/168946
Title: Pemodelan Sentimen Masyarakat Terhadap Kebijakan Pemindahan Ibu Kota Negara di Media Sosial Menggunakan Graph Neural Network (GNN)
Other Titles: Modeling Public Sentiment Towards The Policy Of Moving The National Capital On Social Media Using Graph Neural Network (GNN)
Authors: Sukoco, Heru
Haryanto, Toto
Aminudin, M.Tasurun
Issue Date: 2025
Publisher: IPB University
Abstract: Pemindahan ibu kota negara Indonesia telah resmi ditetapkan melalui Undang-Undang Nomor 3 Tahun 2022 tentang Ibu Kota Negara (IKN) pada masa pemerintahan Presiden Joko Widodo, dengan lokasi baru di Kalimantan Timur. Keputusan strategis ini memicu beragam reaksi di media sosial, baik dukungan maupun penolakan. Untuk memahami persepsi publik secara lebih mendalam, dilakukan analisis sentimen terhadap opini masyarakat di platform Twitter (X) yang dikumpulkan pada periode Januari hingga Agustus 2024. Analisis ini menggunakan pendekatan Graph Neural Network (GNN) yang memungkinkan representasi kalimat dalam bentuk graf, di mana simpul (node) merepresentasikan kata dan sisi (edge) merepresentasikan hubungan antar kata. Hubungan antar kata dibangun melalui struktur dependency tree yang dihasilkan dengan biaffine parser. Dependency tree merupakan representasi struktur sintaksis sebuah kalimat yang menunjukkan hubungan (dependencies) antara kata-kata di dalam kalimat tersebut. Struktur ini menjadi fondasi bagi GNN dalam memodelkan keterkaitan sintaktis dan semantik antar kata dalam kalimat. Sebelum diproses ke model GNN, setiap kata diubah menjadi representasi vektor (word embedding) menggunakan Word2Vec model pra-latih (pre-trained) Google News. Representasi ini membantu menangkap makna leksikal kata-kata dalam konteks yang lebih kaya, sehingga informasi semantik dapat dimanfaatkan secara optimal oleh GNN dalam proses pembelajaran. Model GNN yang digunakan adalah Graph Attention Network (GAT) yang menerapkan mekanisme perhatian (attention mechanism) untuk memungkinkan setiap simpul memberi bobot berbeda pada informasi dari simpul-simpul tetangganya. Arsitektur yang dikembangkan memiliki dua graph attention layer dengan penerapan multi-head attention. Dalam penelitian ini, dirancang tiga konfigurasi model GAT: Model 1 dengan jumlah head pada layer 1 sebanyak 4 dan pada layer 2 sebanyak 2; Model 2 dengan head 8 dan 1; serta Model 3 dengan head 8 dan 8. Hasil evaluasi menunjukkan bahwa ketiga model mencapai tingkat akurasi yang sama, yaitu 85%. Unjuk kerja model terhadap kelas netral dan positif cukup menjanjikan, dengan nilai recall yang relatif tinggi. Meskipun kinerja pada kelas negatif masih lebih rendah dibandingkan dua kelas lainnya, performanya tetap berada pada tingkat yang dapat diterima. Temuan ini mengindikasikan bahwa pendekatan berbasis GAT dengan Word2Vec embedding cukup efektif untuk menangkap sentimen publik, meskipun masih ada ruang untuk meningkatkan sensitivitas model terhadap opini negatif.
The relocation of Indonesia’s capital city was formally enacted through Law Number 3 of 2022 concerning the Capital City (IKN) during the administration of President Joko Widodo, designating East Kalimantan as the new capital. This strategic policy decision has generated diverse reactions across social media platforms, encompassing both support and opposition. To obtain a more comprehensive understanding of public perception, a sentiment analysis was conducted on opinions expressed on the Twitter (X) platform, with data collected between January and August 2024. The analysis employed a Graph Neural Network (GNN) framework, enabling the representation of sentences as graphs in which nodes correspond to words and edges denote syntactic relationships between them. Inter-word relationships were modeled using dependency trees generated via a biaffine parser. A dependency tree provides a representation of a sentence’s syntactic structure by illustrating the dependency relations among its constituent words. This structure serves as the foundation for the GNN in modeling both syntactic and semantic connections within sentences. Prior to being processed by the GNN, each word was transformed into a vector representation (word embedding) using the pre-trained Word2Vec model based on the Google News corpus. This embedding facilitates the capture of lexical semantics in a richer contextual space, thereby enhancing the GNN’s ability to leverage semantic information during the learning process. The GNN model implemented in this study is the Graph Attention Network (GAT), which incorporates an attention mechanism that allows each node to assign varying weights to information from its neighboring nodes. The architecture consists of two graph attention layers employing multi-head attention. Three configurations of the GAT model were examined: Model 1, with four heads in the first layer and two heads in the second; Model 2, with eight heads in the first layer and one in the second; and Model 3, with eight heads in both layers. Evaluation results indicate that all three configurations achieved an identical accuracy rate of 85%. The models demonstrated promising performance for the neutral and positive sentiment classes, with relatively high recall values. Although performance on the negative sentiment class was comparatively lower, it remained within an acceptable range. These findings suggest that the GAT architecture, when combined with Word2Vec embeddings, is effective in capturing public sentiment, while highlighting opportunities for further enhancement in detecting and classifying negative opinions.
URI: http://repository.ipb.ac.id/handle/123456789/168946
Appears in Collections:MT - School of Data Science, Mathematic and Informatics

Files in This Item:
File Description SizeFormat 
cover_G6501211028_f9cbab4616dd45a5b4484be35a6416de.pdfCover2.8 MBAdobe PDFView/Open
fulltext_G6501211028_1f594681515f469e8528b63a1cff3184.pdf
  Restricted Access
Fulltext3.65 MBAdobe PDFView/Open
lampiran_G6501211028_5db061f95b54481aa417fcd1c9a17ea4.pdf
  Restricted Access
Lampiran6.21 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.