Seleksi Fitur Menggunakan Algoritma Gravitational Search Dengan Simulated Annealing (GRAS) Pada Prediksi Customer Churn

Hendro, Hendro (2026) Seleksi Fitur Menggunakan Algoritma Gravitational Search Dengan Simulated Annealing (GRAS) Pada Prediksi Customer Churn. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7025221010-Doctoral.pdf - Accepted Version Restricted to Registered users only Download (4MB) | Request a copy

Abstract

Pelanggan merupakan sumber pendapatan utama bagi perusahaan, sehingga kehilangan pelanggan (customer churn) dapat berdampak signifikan terhadap kinerja bisnis. Dengan berkembangnya teknologi, persaingan antar perusahaan semakin kompetitif, yang mendorong meningkatnya fenomena churn, yaitu perpindahan pelanggan dari satu penyedia layanan ke penyedia lainnya. Oleh karena itu, perusahaan perlu mengidentifikasi potensi churn sejak dini melalui model prediksi yang akurat agar strategi retensi dapat dilakukan secara lebih efektif. Berbagai pendekatan berbasis machine learning telah dikembangkan untuk tujuan ini, namun performa model sangat bergantung pada kualitas fitur yang digunakan. Keberadaan fitur yang tidak relevan tidak hanya menurunkan akurasi, tetapi juga meningkatkan kompleksitas komputasi. Berbagai penelitian telah memanfaatkan teknik feature selection maupun feature weighting untuk mengatasi permasalahan tersebut. Namun, kedua pendekatan tersebut umumnya masih diterapkan secara terpisah sehingga belum mampu memanfaatkan keunggulan keduanya secara bersamaan. Selain itu, metode wrapper berbasis metaheuristik yang digunakan sering menghadapi biaya komputasi yang tinggi serta kecenderungan terjebak pada solusi lokal (local optima). Penelitian ini mengusulkan sebuah pendekatan terpadu yang mengombinasikan seleksi fitur dan pembobotan fitur untuk meningkatkan kinerja model prediksi churn. Pendekatan ini dibangun dengan memanfaatkan Gravitational Search Algorithm (GSA) sebagai metode optimasi untuk menentukan subset fitur yang optimal sekaligus memberikan bobot pada fitur yang terpilih. Lebih lanjut, penelitian ini mengembangkan varian GSA yang ditingkatkan, yaitu GRAS, dengan mengintegrasikan Simulated Annealing guna memperbaiki keseimbangan antara eksplorasi dan eksploitasi selama proses pencarian solusi. Metode yang diusulkan diuji pada beberapa dataset churn dan dibandingkan dengan model baseline serta metode metaheuristik lainnya. Hasil pengujian menunjukkan bahwa pendekatan seleksi fitur dan pembobotan fitur berbasis GSA mampu meningkatkan performa model klasifikasi sekaligus mengurangi jumlah fitur yang digunakan. Selain itu, metode GRAS yang diusulkan mampu mengurangi jumlah fitur hingga 52,68% dan menghasilkan performa yang kompetitif dibandingkan metode metaheuristik lainnya pada berbagai dataset churn. Hasil penelitian ini memberikan kontribusi dalam bentuk pengembangan framework hybrid berbasis GSA untuk seleksi dan pembobotan fitur, serta peningkatan performa melalui integrasi Simulated Annealing dalam algoritma GRAS pada kasus prediksi customer churn.
===================================================================================================================================
Customers are a primary source of revenue for companies, and customer churn can significantly affect business performance. With the advancement of technology, competition among companies has become increasingly intense, leading to a rise in churn, where customers switch from one service provider to another. Therefore, companies need to identify potential churn at an early stage through accurate predictive models in order to implement more effective retention strategies. Various machine learning approaches have been developed for this purpose; however, model performance is highly dependent on the quality of the features used. The presence of irrelevant features not only reduces accuracy but also increases computational complexity. Various machine learning approaches have been developed for this purpose; however, model performance is highly dependent on the quality of the features used. The presence of irrelevant features not only reduces predictive accuracy but also increases computational complexity. Numerous studies have employed feature selection and feature weighting techniques to address this issue. Nevertheless, these approaches are generally applied separately, limiting their ability to fully exploit the advantages of both methods simultaneously. In addition, wrapper-based metaheuristic methods often suffer from high computational costs and a tendency to become trapped in local optima. This study proposes an integrated approach that combines feature selection and feature weighting to improve churn prediction performance. The proposed approach utilizes the Gravitational Search Algorithm (GSA) as an optimization method to determine the optimal subset of features while simultaneously assigning weights to the selected features. Furthermore, this study develops an enhanced variant of GSA, referred to as GRAS, by integrating Simulated Annealing to improve the balance between exploration and exploitation during the search process. The proposed method is evaluated on multiple churn datasets and compared with baseline models as well as other metaheuristic approaches. The experimental results demonstrate that the GSA-based feature selection and feature weighting approaches improve classification performance while reducing the number of features used. Moreover, the proposed GRAS method was able to reduce the number of features by up to 52.68% while achieving competitive performance compared to other metaheuristic methods across various churn datasets. This study contributes by developing a hybrid GSA-based framework for feature selection and feature weighting, as well as enhancing performance through the integration of Simulated Annealing in the GRAS algorithm for customer churn prediction.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Churn Prediction,Machine Learning,Seleksi Fitur,Pembobotan Fitur, Churn Prediction,Machine Learning,Feature Selection,Feature Weighting
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55001-(S3) PhD Thesis (Comp Science)
Depositing User:	Hendro Hendro
Date Deposited:	05 Aug 2026 01:07
Last Modified:	05 Aug 2026 01:07
URI:	http://repository.its.ac.id/id/eprint/143715

Actions (login required)

View Item