# FA-GNN: Filter and Augment Graph Neural Networks for Account Classification in Ethereum

## TLDR
>
> * Nowadays, millions of accounts in Ethereum have been reported to participate in a variety of businesses, and some of them have been found to be involved in illegal behaviors or even cyber-crimes by exploiting the pseudonymous nature of blockchain.
>
> * There is an urgent need for an effective method to conduct account classification and audit transaction behaviors of each account. In this paper, we model the Ethereum transaction records as a transaction network, and the account classification problem is converted to a node classification problem.
>
> * Based on the Ethereum transaction network, we propose a novel framework named Filter and Augment Graph Neural Network (FA-GNN), which can retain the information of important neighbors and augment node features with high-order information. Experimental results demonstrate that our proposed FA-GNN outperforms state-of-the-art methods in Ethereum account classification.

## Research Team
Jieli Liu , Jiatao Zheng, Jiajing Wu , Senior Member, IEEE, and Zibin Zheng , Senior Member, IEEE

## Core Research Question

* What are the performances of our models compared to several state-of-the-art in Ethereum account classifications?

## Background

* **Smart contracts** are self-executing lines of code with the terms of an agreement between buyer and seller automatically verified and executed via a computer network contained therein exists across a distributed, decentralized blockchain network.

* **FA-GNN** is called Filter and Augment Graph Neural Network (FA-GNN), which introduces two key components for Ethereum transaction network embedding learning, namely, neighbor filtering and feature augmentation. By utilizing these two components, FA-GNN can retain information with preference and explore information from the higher order neighborhood. We first describe the general framework of FA-GNN, and then detail the design of FA-GNN.

## Summary

* In this work, we proposed a GNN-based model named FAGNN to tackle the account classification task in Ethereum.

* Specifically, we proposed two key components in our model, namely neighbor filtering and feature augmentation. According to node features and the newly designed neighbor importance measurements in the transaction network, the neighbor filtering component was proposed to identify the influential neighbors and remove redundant neighborhood information. Then enhanced features can be generated for nodes by higherorder information aggregation in the feature augmentation component. Finally, a deep neural network architecture was adopted to learn node representations for downstream tasks.

* Extensive experiments demonstrated the effectiveness of the proposed FAGNN in Ethereum account classification and proved that our model can filter out noised neighbors and explore higher-order information in the Ethereum transaction network.

## Method

The framework of FA-GNN is shown in Fig. 2. As shown in the figure, it contains three main components including neighbor filtering, feature augmentation, and network embedding.
1) Neighbor Filtering: In a transaction network, the relationship between a pair of connected nodes is related to transaction features such as transaction amount and network features such as degree. To remove the redundant information brought by neighbors, we design this component and filter out a certain ratio of neighbors through various preference strategies.

2) Feature Augmentation: Aiming to explore the higher-order information in the network embedding learning process, we propose a feature augmentation component to generate enhanced features for each node by aggregating the higherorder information.

3) Node Representation Learning: With the augmented node features as the input, the node representation learning component learns a low-dimensional representation vector for each node in the network via a two-layered neural network. Subsequently, the obtained representation vectors are used as an input of downstream Ethereum transaction network analysis tasks.


## Results

The first table is FA-GNN model, which improved Precision and Recall compared with state-of-the-art Algorithm (GCN, SGC, GraphSAGE, Cluster-GCN, H2GCN)


The second table is FA-GNN model, which improved Micro-F1 compared with state-of-the-art Algorithm (GCN, SGC, GraphSAGE, Cluster-GCN, H2GCN)


## Conclusion

* In this work, we proposed a GNN-based model named FAGNN to tackle the account classification task in Ethereum. By conducting a network property analysis, we found that there exist great differences between two connected nodes in the Ethereum transaction network in terms of both the node features and node labels.

* Extensive experiments demonstrated the effectiveness of the proposed FA-GNN in Ethereum account classification and proved that our model can filter out noised neighbors and explore higher-order information in the Ethereum transaction network.