Prediksi Churn Nasabah Perbankan dengan Graph Neural Networks dan Explainable AI Berbasis LLM

Hersyaputra, Mohamad Syazimmi (2026) Prediksi Churn Nasabah Perbankan dengan Graph Neural Networks dan Explainable AI Berbasis LLM. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241017-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Churn nasabah merupakan isu strategis dalam perbankan karena berdampak langsung pada laba, nilai bisnis, dan reputasi. Berbagai metode prediksi churn berbasis Machine Learning (ML) seperti mulai dari classifier dasar seperti Decision Trees (DT), hingga ensemble learning seperti Random Forests (RF) dan XGBoost, serta deep learning seperti Artificial Neural Networks (ANN) telah banyak digunakan, namun pendekatan tersebut berbasis data tabular sehingga tidak mampu menangkap relasi laten antar nasabah. Selain itu, banyak model yang telah mencapai performa tinggi bersifat black box dan belum memenuhi kebutuhan transparansi yang penting dalam konteks pengambilan keputusan khsusunya domain berisiko tinggi seperti perbankan. Untuk mengatasi keterbatasan tersebut, penelitian ini mengusulkan pendekatan graph representation learning yang lebih akurat serta komprehensif menggunakan Graph Convolutional Networks (GCN) dan Graph Attention Networks (GAT) sebagai studi komparatif yang diintegrasikan dengan mekanisme Jumping Knowledge (JK) Networks aggregation untuk mengatasi risiko over-smoothing serta meningkatkan kualitas representasi node. Eksperimen dilakukan pada dataset Bank Credit Card Customers (Kaggle × Analyttica Leaps, CC0 Public Domain License) dengan graph construction berbasis weighted cosine similarity yang diberi bobot oleh feature importance RF agar kedekatan antar node mencerminkan kemiripan perilaku nasabah secara relevan. GraphSMOTE diterapkan untuk menangani ketidakseimbangan kelas. Hasil penelitian menunjukkan bahwa integrasi JK Networks menghasilkan peningkatan performa dibandingkan arsitektur konvensionalnya, JK-GCN meningkatkan accuracy sebesar 0,54% dibandingkan GCN konvensional, sedangkan JK-GAT meningkatkan accuracy sebesar 1,08% dibandingkan GAT konvensional. Secara keseluruhan perbandingan, JK-GAT memberikan performa terbaik (accuracy 98,22% dan F1-score 94,48%), sementara JK-GCN mencapai accuracy 97,83% dan F1-score 93,19%, dengan efisiensi komputasi yang lebih baik. Temuan ini menunjukkan bahwa mekanisme agregasi lintas-layer pada JK Networks memperkaya representasi node dan meningkatkan stabilitas. Untuk meningkatkan interpretabilitas, penelitian ini memanfaatkan GNNExplainer untuk mengidentifikasi fitur yang berkontribusi pada prediksi model. Hasil interpretasi tersebut diintegrasikan dengan Large Language Models (LLM) Gemini untuk menghasilkan narasi penjelasan dan rekomendasi retensi berbasis data. Evaluasi 26 praktisi perbankan menunjukkan indeks kualitas narasi sebesar 80,59% (sangat baik), menandakan bahwa integrasi XAI–LLM efektif meningkatkan transparansi.
===================================================================================================================================
Customer churn is a strategic issue in banking because it directly affects revenue, business value, and reputation. Various Machine Learning (ML), based churn prediction methods, ranging from basic classifiers such as Decision Trees (DT), to ensemble learning models such as Random Forests (RF) and XGBoost, as well as deep learning approaches like Artificial Neural Networks (ANN), have been widely used; however, these tabular-based approaches are unable to capture latent relationships among customers. In addition, many high-performing models exhibit black-box behavior and do not meet the transparency requirements essential for decision-making, particularly in high-risk domains such as banking. To address these limitations, this study proposes a more accurate and comprehensive graph representation learning approach using Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) as comparative models, integrated with Jumping Knowledge (JK) Networks aggregation to mitigate over-smoothing and enhance node representation quality. Experiments were conducted on the Bank Credit Card Customers dataset (Kaggle × Analyttica Leaps, CC0 Public Domain License) using graph construction based on weighted cosine similarity, where weights were assigned using RF feature importance so that node proximity reflects relevant similarities in customer behavior. GraphSMOTE was applied to address class imbalance. The results show that the integration of JK Networks improves performance compared to their conventional architectures; JK-GCN increases accuracy by 0.54% over conventional GCN, while JK-GAT increases accuracy by 1.08% over conventional GAT. Overall, JK-GAT provides the best performance (98.22% accuracy and 94.48% F1-score), while JK-GCN achieves 97.83% accuracy and 93.19% F1-score with better computational efficiency. These findings indicate that cross-layer aggregation mechanisms in JK Networks enrich node representations and improve stability. To enhance interpretability, this study employs GNNExplainer to identify features contributing to the model’s predictions. The interpretation results are integrated with the Gemini Large Language Models (LLM) to produce natural-language explanations and data-driven retention recommendations. An evaluation involving 26 banking practitioners shows a narrative quality index of 80.59% (very good), indicating that the XAI–LLM integration effectively improves transparency.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Prediksi churn nasabah perbankan, Graph Neural Networks, Jumping Knowledge Networks, GraphSMOTE, LLM, Explainable AI, Bank customer churn prediction, Graph Neural Networks, Jumping Knowledge Networks, GraphSMOTE, LLM, Explainable AI
Subjects:	T Technology > T Technology (General)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Mohamad Syazimmi Hersyaputra
Date Deposited:	29 Jan 2026 07:45
Last Modified:	29 Jan 2026 07:45
URI:	http://repository.its.ac.id/id/eprint/131036

Actions (login required)

View Item