Deteksi Penipuan Transaksi E-Commerce Menggunakan Metode GNN Dan Pendekatan Machine Learning

Kurniadi, Nathaniel Ryo (2026) Deteksi Penipuan Transaksi E-Commerce Menggunakan Metode GNN Dan Pendekatan Machine Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221019-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Pertumbuhan e-commerce yang pesat diikuti oleh meningkatnya kasus penipuan transaksi digital. Sistem deteksi berbasis aturan deterministik tidak lagi efektif karena hanya mampu mengenali pola yang telah ditetapkan sebelumnya, sehingga sulit beradaptasi terhadap modus penipuan yang terus berkembang. Pendekatan ini mengandalkan aturan yang ditentukan secara manual, seperti ambang nominal transaksi, aturan frekuensi transaksi, dan daftar hitam kartu atau pedagang. Pendekatan machine learning menawarkan solusi yang lebih adaptif karena mampu mempelajari pola dari data transaksi historis, seperti umur akun, frekuensi transaksi, dan rata-rata nilai transaksi pengguna. Namun, pendekatan ini masih memiliki keterbatasan karena pelaku terus mengubah pola dan taktik serangannya untuk menghindari sistem deteksi. Pendekatan dengan Graph Neural Network (GNN) kemudian ditelusuri karena mampu memodelkan hubungan antar entitas, seperti keterkaitan akun dan kesamaan metadata pembayaran, yang sulit ditangkap oleh metode tabular, namun penerapannya secara mandiri menuntut komputasi yang tinggi. Dengan demikian, bergantung sepenuhnya pada satu pendekatan saja memaksa penerimaan atas keterbatasan bawaannya: metode tabular unggul dalam efisiensi komputasi namun lemah dalam menangkap keterkaitan relasional antar entitas dan rentan terhadap perubahan pola serta taktik pelaku, sedangkan GNN unggul dalam pemodelan relasional namun mahal secara komputasi sehingga kurang efisien untuk layanan yang sensitif terhadap latensi. Trade-off inilah yang mendorong perlunya mengintegrasikan kedua pendekatan agar kelebihan masing-masing dapat saling melengkapi sekaligus menutupi keterbatasannya. Tugas Akhir ini membandingkan pendekatan tabular dan berbasis graf sekaligus mengeksplorasi penggabungannya melalui feature embedding menggunakan Heterogeneous Graph Transformer (HGT). Evaluasi dilakukan melalui lima skenario yang terbagi menjadi dua lingkup. Pada lingkup pemodelan, dilakukan tiga skenario: (1) machine learning tabular sebagai baseline menggunakan 26 fitur tabular, (2) HGT end-to-end sebagai pengklasifikasi mandiri pada graf heterogen, dan (3) pendekatan hibrida yang menggabungkan 26 fitur tabular dengan 64 dimensi embedding HGT. Pada lingkup service, dilakukan dua skenario pengujian, yaitu pengujian fungsionalitas dengan blackbox testing dan pengujian usability dengan System Usability Scale (SUS). Evaluasi menggunakan dataset sintetis E-Commerce Fraud Detection Dataset yang terdiri dari 299.695 transaksi. Model terbaik kemudian diimplementasikan sebagai service deteksi penipuan bernama Smart Intelligence for Guarding and Analyzing Payments (SIGAP). Hasil Tugas Akhir menunjukkan bahwa model terbaik adalah Gradient Boosting pada pendekatan hibrida (Skenario 3) dengan F1-score 0,787, accuracy 0,991, precision 0,861, recall 0,725, MCC 0,786, dan PR-AUC 0,829. Pengujian fungsionalitas service SIGAP berhasil pada seluruh kasus uji, dan pengujian usability memperoleh skor SUS 67,75 (grade D, adjective rating OK). Temuan ini menunjukkan bahwa pemanfaatan GNN sebagai feature extractor mampu memperkaya representasi fitur tabular melalui informasi relasional antar entitas. Implementasi dan pengujian service SIGAP juga menunjukkan bahwa pendekatan hibrida berbasis GNN dan machine learning layak diterapkan untuk mendukung deteksi penipuan transaksi e-commerce, dengan pengembangan lebih lanjut pada aspek panduan penggunaan untuk meningkatkan kemudahan operasional bagi pengguna.
=================================================================================================================================================
The rapid growth of e-commerce has been accompanied by a rising number of digital transaction fraud cases. Deterministic rule-based detection systems are no longer effective because they can only recognize predefined patterns, making them difficult to adapt to continuously evolving fraud schemes. These systems rely on manually defined rules, such as transaction amount thresholds, transaction frequency rules, and blacklists of cards or merchants. Machine learning approaches offer a more adaptive solution because they can learn patterns from historical transaction data, such as account age, transaction frequency, and a user's average transaction value. However, this approach still has limitations, as fraudsters continually change their patterns and attack tactics to evade detection systems. The Graph Neural Network (GNN) approach was then explored because it can model relationships between entities, such as account linkages and the similarity of payment metadata, which are difficult for tabular methods to capture; however, its standalone implementation demands high computational cost. Consequently, relying entirely on a single approach forces the acceptance of its inherent limitations: tabular methods excel in computational efficiency but are weak at capturing relational connections between entities and remain vulnerable to shifts in fraudsters' patterns and tactics, whereas GNNs excel at relational modeling but are computationally expensive, making them less efficient for latency-sensitive services. This trade-off motivates the need to integrate both approaches so that the strengths of each can complement one another while offsetting their respective limitations. This Final Project compares tabular and graph-based approaches while also exploring their integration through feature embedding using a Heterogeneous Graph Transformer (HGT). The evaluation consists of five scenarios divided into two scopes. In the modeling scope, three scenarios are examined: (1) tabular machine learning as a baseline using 26 tabular features, (2) an end-to-end HGT model as a standalone classifier on a heterogeneous graph, and (3) a hybrid approach combining 26 tabular features with 64-dimensional HGT embeddings. In the service scope, two evaluation scenarios are conducted, namely functionality testing using black box testing and usability testing using the System Usability Scale (SUS). The evaluation utilizes the synthetic E-Commerce Fraud Detection Dataset containing 299,695 transactions. The best performing model is subsequently implemented as a fraud detection service named Smart Intelligence for Guarding and Analyzing Payments (SIGAP). The results show that the best-performing model is Gradient Boosting under the hybrid approach (Scenario 3), achieving an F1-score of 0.787, accuracy of 0.991, precision of 0.861, recall of 0.725, MCC of 0.786, and PR-AUC of 0.829. Functionality testing of the SIGAP service passed all test cases, while usability testing achieved an SUS score of 67.75 (Grade D, adjective rating: OK). These findings indicate that utilizing GNN as a feature extractor can enrich tabular feature representations through relational information among entities. The implementation and evaluation of the SIGAP service also demonstrate that a hybrid approach combining GNN and machine learning is feasible for supporting e-commerce transaction fraud detection. Further improvements, particularly in user guidance and documentation, are recommended to enhance operational usability for end users.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Penipuan, E-Commerce, Graph Neural Network, Machine Learning, Feature Embedding
Subjects:	T Technology > T Technology (General) > T174.5 Technology--Risk assessment. T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models. T Technology > T Technology (General) > T58.6 Management information systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Nathaniel Ryo Kurniadi
Date Deposited:	17 Jul 2026 05:48
Last Modified:	17 Jul 2026 05:48
URI:	http://repository.its.ac.id/id/eprint/135278

Actions (login required)

View Item