Explorer›Research Paper›Computer Science
Research PaperResearchia:202609.10097

fraud detection eliptical bitcoin dataset

Anonymous

Abstract

This research presents a chronological evaluation of a hybrid quantum–classical Graph Neural Network (GNN) for illicit Bitcoin transaction detection using the Elliptic Bitcoin Transaction Dataset. The study investigates whether incorporating a small Variational Quantum Circuit (VQC) into a graph-based learning pipeline improves illicit transaction classification compared with a classical Graph Convolutional Network (GCN).

The Elliptic dataset contains 203,769 Bitcoin transactions connected by 234,355 directed transaction-flow edges across 49 time steps, with 165 model input features per transaction. For supervised learning, licit and illicit transactions were used, while transactions with unknown labels were excluded. To reduce temporal leakage and provide a more realistic evaluation, the data were chronologically divided into training (time steps 1–34), validation (35–41), and held-out testing (42–49). Feature standardization was fitted only on the training-period data.

Two models were evaluated. The classical baseline was a two-layer GCN with 5,378 trainable parameters. The hybrid quantum–classical model combined graph convolution with a 4-qubit, 2-layer Variational Quantum Circuit and contained 6,518 trainable parameters. Due to the computational cost of differentiable quantum simulation on CPU, the quantum training stage was restricted to 256 selected training nodes.

On the held-out test period, the classical GCN achieved an AUPRC of 0.145466, ROC-AUC of 0.771520, F1-score of 0.146276, precision of 0.080482, and recall of 0.801471. The hybrid quantum model achieved an AUPRC of 0.038754, ROC-AUC of 0.448034, F1-score of 0.066192, precision of 0.035916, and recall of 0.421569.

The results show that the classical GCN outperformed the evaluated hybrid quantum–classical configuration on all five held-out test metrics. Therefore, no quantum performance advantage was demonstrated in this experimental setting. However, the findings are specific to the tested 4-qubit VQC configuration and do not establish that quantum graph learning is generally inferior to classical graph learning. The study highlights the importance of chronological evaluation, class-imbalance-aware metrics, transparent quantum benchmarking, and careful interpretation of quantum machine learning results in financial transaction analysis.

Submitted: September 10, 2026Subjects: Computer Science; Research Paper

Description / Details

<div><div><span style="font-size: 17.6px;">Research Title:</span></div><div><span style="font-size: 17.6px;">Chronological Evaluation of a Hybrid Quantum–Classical Graph Neural Network for Illicit Bitcoin Transaction Detection</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Research Problem:</span></div><div><span style="font-size: 17.6px;">Financial transaction networks are naturally represented as graphs because transactions are connected through transaction flows. Graph Neural Networks can exploit these relationships for illicit transaction detection. At the same time, quantum machine learning has been proposed as a potential extension to classical machine learning. However, quantum components should not be assumed to provide an advantage without direct experimental comparison. This research therefore evaluates a classical GCN against a hybrid GCN–VQC model under a chronological Bitcoin transaction setting.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Dataset:</span></div><div><span style="font-size: 17.6px;">Elliptic Bitcoin Transaction Dataset</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Dataset Characteristics:</span></div><div><span style="font-size: 17.6px;">• Total transactions/nodes: 203,769</span></div><div><span style="font-size: 17.6px;">• Directed transaction-flow edges: 234,355</span></div><div><span style="font-size: 17.6px;">• Model input features per transaction: 165</span></div><div><span style="font-size: 17.6px;">• Total time steps: 49</span></div><div><span style="font-size: 17.6px;">• Classes used for supervised learning: Licit and Illicit</span></div><div><span style="font-size: 17.6px;">• Unknown-labelled transactions: Excluded from supervised learning/evaluation</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Input to the Model:</span></div><div><span style="font-size: 17.6px;">The input consists of 165-dimensional feature vectors for Bitcoin transaction nodes together with the graph connectivity representing transaction flow between transactions. Therefore, the GCN learns from both individual transaction characteristics and relational information from connected transactions.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Preprocessing:</span></div><div><span style="font-size: 17.6px;">• Transaction ID and time information were excluded from model input features.</span></div><div><span style="font-size: 17.6px;">• Unknown-labelled transactions were excluded from the supervised classification task.</span></div><div><span style="font-size: 17.6px;">• StandardScaler was fitted using training-period data only.</span></div><div><span style="font-size: 17.6px;">• The fitted scaler was subsequently applied to validation and test data.</span></div><div><span style="font-size: 17.6px;">• Separate graph objects were constructed for the chronological periods.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Chronological Data Split:</span></div><div><span style="font-size: 17.6px;">Training:</span></div><div><span style="font-size: 17.6px;">• Time steps: 1–34</span></div><div><span style="font-size: 17.6px;">• Nodes: 29,894</span></div><div><span style="font-size: 17.6px;">• Edges: 22,898</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Validation:</span></div><div><span style="font-size: 17.6px;">• Time steps: 35–41</span></div><div><span style="font-size: 17.6px;">• Nodes: 7,829</span></div><div><span style="font-size: 17.6px;">• Edges: 6,509</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Held-out Test:</span></div><div><span style="font-size: 17.6px;">• Time steps: 42–49</span></div><div><span style="font-size: 17.6px;">• Nodes: 8,841</span></div><div><span style="font-size: 17.6px;">• Edges: 7,217</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Training Class Distribution:</span></div><div><span style="font-size: 17.6px;">• Licit transactions: 26,432</span></div><div><span style="font-size: 17.6px;">• Illicit transactions: 3,462</span></div><div><span style="font-size: 17.6px;">• Class weighting was used because of the class imbalance.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Classical Model:</span></div><div><span style="font-size: 17.6px;">A two-layer Graph Convolutional Network was implemented as the classical baseline.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Architecture:</span></div><div><span style="font-size: 17.6px;">165 input features</span></div><div><span style="font-size: 17.6px;">→ GCN layer (165 → 32)</span></div><div><span style="font-size: 17.6px;">→ ReLU</span></div><div><span style="font-size: 17.6px;">→ GCN layer (32 → 2)</span></div><div><span style="font-size: 17.6px;">→ Licit/Illicit prediction</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Trainable Parameters:</span></div><div><span style="font-size: 17.6px;">5,378</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Hybrid Quantum–Classical Model:</span></div><div><span style="font-size: 17.6px;">The hybrid model combines graph-based feature extraction with a Variational Quantum Circuit.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Architecture:</span></div><div><span style="font-size: 17.6px;">165 input features</span></div><div><span style="font-size: 17.6px;">→ GCN layer (165 → 32)</span></div><div><span style="font-size: 17.6px;">→ ReLU</span></div><div><span style="font-size: 17.6px;">→ graph representation</span></div><div><span style="font-size: 17.6px;">→ linear projection (32 → 4)</span></div><div><span style="font-size: 17.6px;">→ 4-qubit Variational Quantum Circuit</span></div><div><span style="font-size: 17.6px;">→ 2 variational layers</span></div><div><span style="font-size: 17.6px;">→ quantum expectation-value measurements</span></div><div><span style="font-size: 17.6px;">→ classical output classifier</span></div><div><span style="font-size: 17.6px;">→ Licit/Illicit prediction</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Quantum Configuration:</span></div><div><span style="font-size: 17.6px;">• Number of qubits: 4</span></div><div><span style="font-size: 17.6px;">• Variational layers: 2</span></div><div><span style="font-size: 17.6px;">• VQC trainable parameters: 8</span></div><div><span style="font-size: 17.6px;">• Hybrid model total trainable parameters: 6,518</span></div><div><span style="font-size: 17.6px;">• Quantum execution: simulator</span></div><div><span style="font-size: 17.6px;">• Random seed: 42</span></div><div><span style="font-size: 17.6px;">• Differentiable quantum training: 256 selected training nodes because of CPU simulation cost</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Evaluation Metrics:</span></div><div><span style="font-size: 17.6px;">Because the dataset is highly imbalanced, accuracy alone was not used as the main indicator. The models were evaluated using:</span></div><div><span style="font-size: 17.6px;">• AUPRC</span></div><div><span style="font-size: 17.6px;">• ROC-AUC</span></div><div><span style="font-size: 17.6px;">• F1-score</span></div><div><span style="font-size: 17.6px;">• Precision</span></div><div><span style="font-size: 17.6px;">• Recall</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Held-Out Test Results:</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Classical GCN:</span></div><div><span style="font-size: 17.6px;">• AUPRC: 0.145466</span></div><div><span style="font-size: 17.6px;">• ROC-AUC: 0.771520</span></div><div><span style="font-size: 17.6px;">• F1-score: 0.146276</span></div><div><span style="font-size: 17.6px;">• Precision: 0.080482</span></div><div><span style="font-size: 17.6px;">• Recall: 0.801471</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Hybrid Quantum Model:</span></div><div><span style="font-size: 17.6px;">• AUPRC: 0.038754</span></div><div><span style="font-size: 17.6px;">• ROC-AUC: 0.448034</span></div><div><span style="font-size: 17.6px;">• F1-score: 0.066192</span></div><div><span style="font-size: 17.6px;">• Precision: 0.035916</span></div><div><span style="font-size: 17.6px;">• Recall: 0.421569</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Main Findings:</span></div><div><span style="font-size: 17.6px;">1. The classical GCN outperformed the tested hybrid quantum model on every reported held-out test metric.</span></div><div><span style="font-size: 17.6px;">2. The classical GCN achieved substantially higher AUPRC and ROC-AUC than the hybrid model.</span></div><div><span style="font-size: 17.6px;">3. Classical recall reached 0.801471, showing that it identified a relatively large proportion of illicit transactions at the fixed 0.5 decision threshold.</span></div><div><span style="font-size: 17.6px;">4. Classical precision remained low at 0.080482, indicating that false positives remain an important problem.</span></div><div><span style="font-size: 17.6px;">5. The tested 4-qubit VQC did not improve predictive performance over the classical GCN.</span></div><div><span style="font-size: 17.6px;">6. No quantum performance advantage was demonstrated in this experimental configuration.</span></div><div><span style="font-size: 17.6px;">7. Quantum simulation introduced a practical computational limitation, requiring the differentiable quantum stage to be restricted to 256 training nodes.</span></div><div><span style="font-size: 17.6px;">8. Chronological splitting and training-only feature scaling reduced obvious future-information leakage compared with a random-split approach.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">What the Research Establishes:</span></div><div><span style="font-size: 17.6px;">The experiment establishes that, under the implemented chronological Elliptic Bitcoin transaction setting, the classical GCN provided better held-out predictive performance than the evaluated 4-qubit hybrid GCN–VQC configuration. The result provides empirical evidence that adding a quantum component does not automatically improve a graph-based financial transaction classifier.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">The research does NOT establish that all classical GNNs are superior to all quantum GNNs. The conclusion applies only to the architecture, dataset, training conditions, and quantum configuration evaluated in this experiment.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Limitations:</span></div><div><span style="font-size: 17.6px;">• Only one random seed (42) was evaluated.</span></div><div><span style="font-size: 17.6px;">• Classical and hybrid models were not parameter matched (5,378 vs 6,518 parameters).</span></div><div><span style="font-size: 17.6px;">• Differentiable quantum training was restricted to 256 selected training nodes.</span></div><div><span style="font-size: 17.6px;">• Quantum execution used CPU simulation rather than real quantum hardware.</span></div><div><span style="font-size: 17.6px;">• Only one 4-qubit, 2-layer VQC architecture was evaluated.</span></div><div><span style="font-size: 17.6px;">• The experiment used one dataset.</span></div><div><span style="font-size: 17.6px;">• A fixed classification threshold of 0.5 was used.</span></div><div><span style="font-size: 17.6px;">• Archived quantum validation records contained an inconsistency; therefore, disputed quantum validation values were excluded from the main quantitative claims.</span></div><div><span style="font-size: 17.6px;">• Stronger temporal GNN architectures such as EvolveGCN or TGN were not compared under identical preprocessing.</span></div><div><span style="font-size: 17.6px;">• Federated Learning and FedAvg were NOT implemented in this experiment; therefore, no federated-learning, privacy-preserving, or cross-institutional performance claim is made.</span></div><div><span style="font-size: 17.6px;"><br></span></div><div><span style="font-size: 17.6px;">Final Conclusion:</span></div><div><span style="font-size: 17.6px;">Under chronological and class-imbalance-aware evaluation on the Elliptic Bitcoin transaction graph, the classical GCN performed better than the tested hybrid quantum–classical model across AUPRC, ROC-AUC, F1-score, precision, and recall. The evaluated quantum component did not provide a performance advantage in this setting. These findings demonstrate the importance of experimentally validating proposed quantum advantages rather than assuming that the addition of a quantum circuit will necessarily improve financial fraud or illicit-transaction detection.</span></div></div>

Please sign in to join the discussion.

No comments yet. Be the first to share your thoughts!

Access Paper
Submission Info
Date:
Sep 10, 2026
Topic:
Research Paper
Area:
Computer Science
Comments:
0
Bookmark
fraud detection eliptical bitcoin dataset | Researchia