1 Introduction
-
To present a model in a novel way that considers the graph structure and temporal sequence of Elliptic data to predict illicit transactions that belong to illegal services in Bitcoin blockchain network.
-
To perform active learning on Bitcoin blockchain data that mimics a real-life situation, since Bitcoin blockchain is a massively growing technology and its data is time-consuming to label.
2 Overview of Related Work
3 Model Uncertainty: MC-Dropout Versus MC-AA
3.1 MC-Dropout: Monte-Carlo Dropout
3.2 MC-AA: Monte-Carlo Based Adversarial Attack
4 Acquisition Functions for Active Learning
4.1 BALD: Bayesian Active Learning by Disagreement
4.2 Entropy
4.3 Variation Ratios
4.4 Mean Standard Deviation
4.5 Random Sampling: Baseline Model
5 Methods
5.1 Dataset Description
5.2 Temporal Modelling
5.2.1 Long Short-Term Memory (LSTM)
5.2.2 Topology Adaptive Graph Convolutional Network: TAGCN
5.2.3 Overall Model: Temporal-GCN
6 Experiments
6.1 Experimental Setup
Model | % Accuracy | % Precision | % Recall | % F1 Score |
---|---|---|---|---|
Temporal-GCN | 97.7 | 92.7 | 71.3 | 80.6 |
GCN + MLP [5] | 97.4 | 89.9 | 67.8 | 77.3 |
Evolve-GCN [3] | 96.8 | 85.0 | 62.4 | 72.0 |
Skip-GCN [3] | 96.6 | 81.2 | 62.3 | 70.5 |
Random Forest [3] | 96.6 | 80.3 | 61.1 | 69.4 |
GCN [3] | 96.1 | 81.2 | 51.2 | 62.8 |
MLP [3] | 95.8 | 63.7 | 66.2 | 64.9 |
Logistic regression [3] | 92.0 | 34.8 | 66.8 | 45.7 |
6.2 Active Learning
Acquisition function | Runtime (mins) using MC-AA | Runtime (mins) using MC-dropout |
---|---|---|
BALD | tMC−AA = 28.07 | tMC − dropout = 27.1 |
Entropy | tMC−AA = 28.9 | tMC − dropout = 27.3 |
Variation ratio | tMC−AA = 28.9 | tMC − dropout = 28.3 |
Mean STD | tMC−AA = 28.68 | tMC − dropout = 27.09 |
7 Results and Discussion
7.1 Performance of Temporal-GCN
7.2 Evaluation of Active Learning Frameworks
7.3 Wilcoxon Hypothesis Test
Acquisition function | Wilcoxon (MC-AA, random sampling) | Wilcoxon (MC-dropout, random sampling) |
---|---|---|
BALD | 0.0009 | 0.0217 |
Entropy | 0.2552 | 0.309 |
Variation ratio | 0.266 | 0.0071 |
Mean STD | 0.5507 | 0.0112 |
8 Ablation Study
Model number | Layer-1 | Layer-2 | Layer-3 | % Accuracy |
---|---|---|---|---|
Model-0 (ours) | LSTM | TAGConv | TAGConv | 97.77 |
Model-1 | Linear | TAGConv | TAGConv | 97.66 |
Model-2 | LSTM | Linear | TAGConv | 97.76 |
Model-3 | LSTM | TAGConv | Linear | 97.57 |
Model-4 | LSTM | Linear | Linear | 97.44 |
Model-5 | Linear | Linear | Linear | 97.51 |
Model-6 | LSTM | TAGConv | None | 97.63 |
Model-7 | TAGConv | TAGConv | None | 96.65 |
TAGCN K-hops | % Accuracy |
---|---|
K = 3 | 97.77 |
K = 2 | 97.71 |
K = 1 | 97.75 |