Identifikasi Faktor Kerawanan Pangan Rumah Tangga di Maluku melalui Pendekatan Heterogeneous GraphSAGE

Abstract

Kerawanan pangan merupakan permasalahan yang dipengaruhi oleh karakteristik rumah tangga dan juga kondisi wilayah tempat rumah tangga berada. Pendekatan machine learning berbasis data tabular umumnya mengasumsikan setiap rumah tangga bersifat independen sehingga belum mampu memanfaatkan hubungan antarentitas. Penelitian bertujuan menganalisis pengaruh variasi konstruksi graf heterogen terhadap model GraphSAGE, menentukan konstruksi graf terbaik, serta mengidentifikasi fitur penting terhadap kerawanan pangan rumah tangga di Maluku. Data yang digunakan berasal dari Survei Sosial Ekonomi Nasional periode Maret 2024 mencakup 11.381 rumah tangga di Maluku. Empat konstruksi graf heterogen dibangun berdasarkan hubungan administratif rumah tangga, kedekatan spasial antarkabupaten, dan kemiripan karakteristik hunian antarrumah tangga. Model GraphSAGE diimplementasikan pada graf heterogen dan dievaluasi menggunakan metrik balanced accuracy. GNNExplainer digunakan untuk menghitung skor atribusi edge dan Integrated Gradients digunakan untuk mengidentifikasi fitur penting. Hasil penelitian menunjukan konstruksi graf heterogen memengaruhi cara model GraphSAGE memanfaatkan informasi dari setiap tipe edge dalam mengklasifikasikan status kerawanan pangan. Konstruksi yang menggabungkan edge administratif dan edge spasial menghasilkan performa terbaik, sedangkan penambahan edge kemiripan pada konstruksi graf yang menggabungkan seluruh tipe edge tidak meningkatkan performa model. Fitur lapangan pekerjaan kepala rumah tangga, pengeluaran per kapita, dan pendidikan kepala rumah tangga merupakan fitur penting dalam klasifikasi kerawanan pangan di Maluku.

Food insecurity is influenced by both household characteristics and the conditions of the areas in which households are located. Conventional tabular machine learning approaches generally assume that households are independent, limiting their ability to capture relationships among entities. This study analyzes the effect of heterogeneous graph construction on GraphSAGE performance, determines the optimal graph construction, and identifies important features associated with household food insecurity in Maluku. Data were obtained from the March 2024 National Socioeconomic Survey (SUSENAS), comprising 11,381 households. Four heterogeneous graph constructions were developed based on administrative relationships, spatial proximity among regencies/municipalities, and similarities in housing characteristics. GraphSAGE was implemented on the heterogeneous graphs and evaluated using balanced accuracy. GNNExplainer computed edge attribution scores, while Integrated Gradients identified important features. Results indicate that heterogeneous graph construction influences how GraphSAGE utilizes information from different edge types. The graph combining administrative and spatial edges achieved the best performance, whereas adding similarity edges to the graph containing all edge types did not improve performance. The occupation and educational attainment of the household head, along with per capita expenditure, were identified as the most important features for classifying household food insecurity in Maluku.