Data-Warehouse-Enhanced Machine Learning Framework for Multi-Perspective Fraud Detection in Multi-Stakeholder E-Commerce Transactions - Research Paper

Aug 4, 2026

Upload by Ophelia

E-commerce fraud increasingly depends on multiple stakeholders—buyers, sellers, logistics providers, and payment gateways—creating cross-entity fraud patterns that traditional detection struggles to identify. Although machine learning improves predictive performance, fragmented silo datasets prevent capturing multi-perspective behavioral signals and limit feature engineering. This paper presents a data-warehouse-enhanced framework consolidating heterogeneous stakeholder data into a unified environment for scalable fraud modeling. It combines Random Forest, LSTM, Graph Neural Networks, and One-Class SVM, improving accuracy, lowering false positives, and strengthening temporal and relational pattern discovery against non-warehouse and single-perspective baselines.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Abstract

Introduction

Multi-stakeholder fraud challenges

From rule-based methods to ML classifiers

Temporal modeling with LSTM

Relational modeling with GNNs

Unsupervised anomaly detection with One-Class SVM

Data fragmentation and the warehouse gap

Proposed warehouse-enhanced framework capabilities

Data-Warehouse-Enhanced Machine Learning Framework for Multi-Perspective Fraud Detection in Multi-Stakeholder E-Commerce Transactions Naga Charan Nandigama Independent Researcher, Tampa, Florida, USA ABSTRACT E-commerce fraud has grown increasingly complex due to the involvement of multiple stakeholders—buyers, sellers, logistics providers, and payment gateways—leading to sophisticated cross-entity fraud patterns that traditional detection systems struggle to identify. While modern machine-learning techniques offer improved predictive capabilities, their effectiveness is often limited by fragmented, siloed datasets that fail to capture multi-perspective behavioural signals. This paper proposes a Data-Warehouse-Enhanced Machine Learning Framework that consolidates heterogeneous stakeholder data into a unified analytical environment, enabling richer feature engineering and scalable fraud modeling. The framework integrates multiple machine-learning algorithms—Random Forest (RF) for robust supervised classification, Long Short-Term Memory (LSTM) networks for temporal transaction modeling, Graph Neural Networks (GNNs) for capturing relational and cross-stakeholder dependencies, and One-Class SVM for anomaly detection under extreme class imbalance. Experimental evaluations demonstrate that the warehouse-enhanced multi-perspective learning approach significantly improves fraud-classification accuracy, reduces false positives, and enhances temporal and relational pattern discovery compared to non-warehouse and single-perspective baselines. The proposed system provides an effective and scalable foundation for next-generation fraud detection in multi-stakeholder e-commerce ecosystems. Keywords: E-commerce fraud detection, data warehousing, machine learning, multi-stakeholder analytics, multiperspective modeling, anomaly detection, big data architecture. I. INTRODUCTION E-commerce ecosystems today operate as complex digital marketplaces involving multiple interacting stakeholders—buyers, sellers, logistics partners, payment gateways, warehouses, and customer-support entities. This interconnected transactional landscape creates both rich behavioral signals and expanded vulnerability surfaces for fraudulent activities such as synthetic identity fraud, coordinated seller–buyer collusion, triangulation fraud, refund manipulation, false shipment claims, and device/IP spoofing [1]–[4] . As fraudulent behaviors evolve in sophistication, traditional rule-based or isolated analytic systems fail to capture the multi-dimensional interactions across stakeholders. Between 1999 and 2010, fraud mitigation predominantly relied on static rules, regression models, and threshold-based anomaly indicators [1], [2] . While interpretable, these approaches lacked adaptability to changing fraud patterns and did not incorporate multichannel data. The acceleration of data availability and machine-learning advancements shifted the fraud detection paradigm significantly. Algorithms such as Random Forest (RF) emerged as robust classifiers capable of handling heterogeneous features and nonlinear interactions [5]–[8] . RF models are particularly useful in e-commerce systems due to their ability to manage missing values, mixed data types, and high-dimensional behavior logs. However, as fraud strategies increasingly exploit temporal patterns—such as repeated micro-transactions, rapid cart abandonment, abnormal login sequences, or bursty refund claims—researchers adopted Long ShortTerm Memory (LSTM) networks to capture sequential and time-dependent signals [7], [10], [13] . LSTM models effectively learn temporal anomalies and transaction evolution, yet they remain limited when fraud involves networks of interacting entities across buyers, sellers, and logistics agents. To capture these relational dependencies, Graph Neural Networks (GNNs) have recently gained prominence for fraud detection in multi-stakeholder environments. In ecommerce ecosystems wh

FAQs

Why do traditional fraud detection systems struggle in multi-stakeholder e-commerce?

Fraud patterns span multiple entities, but conventional approaches rely on isolated or rule-based signals that cannot capture cross-stakeholder interactions and evolving behaviors across the ecosystem.

How does the framework use data warehousing to improve machine learning performance?

It consolidates heterogeneous stakeholder datasets into a unified structured analytical environment, enabling multi-perspective feature engineering, cross-entity histories, and consistent data refresh for model training and evaluation.

What roles do the different machine learning models play in the proposed system?

Random Forest performs supervised classification with robust handling of mixed and high-dimensional features, LSTM models temporal transaction sequences, GNNs capture relational dependencies and collusion patterns, and One-Class SVM supports anomaly detection under extreme class imbalance.

Aug 4, 2026

Upload by Ophelia