Source-linked AI summary
Modeling and Understanding Ethereum Transaction Records via a Complex Network Approach
Dan Lin, Jiajing Wu, Qi Yuan, Zibin Zheng
TL;DR
Ethereum transaction records remain relatively unexplored, and existing graph-representation methods omit temporal and multiplicity attributes. This paper models transactions as a Temporal Weighted Multidigraph and learns node representations using temporal walks with flexible biases. Temporal information and edge multiplicity are important for transaction-network analysis, and the proposed modeling and biased walks perform better in the reported link-prediction comparisons.
Problem
Ethereum transaction analysis is relatively unexplored, and existing graph-representation methods ignore temporal and multiplicity attributes of financial transaction networks.
Source: §I · p. 1Method
The paper models Ethereum transactions as a Temporal Weighted Multidigraph and learns node representations through feasible temporal walks with flexible time- and amount-based sampling biases.
Source: §I · p. 1, §V · pp. 3–4Results
Temporal Weighted Multidigraph modeling without bias overwhelmingly outperforms traditional directed-graph modeling, while temporal walks biased by time and amount outperform unbiased temporal walks on temporal directed link prediction.
Source: §VI · p. 4Takeaways & Limitations
The reported comparisons indicate that temporal information and edge multiplicity matter for analyzing and understanding financial transaction networks.
Source: §VI · p. 5The temporal-random-walk graph representation is not inductive and cannot obtain representations for newly added nodes.
Source: §VII · p. 5
Abstract
from arXiv · showhide
As the largest public blockchain-based platform supporting smart contracts, Ethereum has accumulated a large number of user transaction records since its debut in 2014. Analysis of Ethereum transaction records, however, is still relatively unexplored till now. Modeling the transaction records as a static simple graph, existing methods are unable to accurately characterize the temporal and multiplex features of the edges. In this brief, we first model the Ethereum transaction records as a complex network by incorporating time and amount features of the transactions, and then design several flexible temporal walk strategies for random-walk based graph representation of this large-scale network. Experiments of temporal link prediction on real Ethereum data demonstrate that temporal information and multiplicity characteristic of edges are indispensable for accurate modeling and understanding of Ethereum transaction networks.
I. INTRODUCTION
Ethereum transaction analysis is relatively unexplored, and existing simple-graph models omit transaction-level temporal and multiplicity features. This brief models Ethereum records as a temporal, weighted multidigraph and evaluates random-walk-based graph representations on link prediction.
I. INTRODUCTION
- Simple-graph models merge multiple transactions and cannot adequately characterize transaction networks’ temporal and multiplicity features.Each Ethereum edge represents a particular Ether transaction with direction, amount, and timestamp.Source: §I · p. 1
- Graph representation methods largely target social networks, while existing methods ignore temporal and multiplicity attributes needed for financial transaction networks.The interaction between Ethereum users and smart contracts remains relatively unexplored.Source: §I · p. 1
- The brief models Ethereum transaction records as a Temporal Weighted Multidigraph and proposes graph sampling and representation methods that incorporate temporal and weighted information.Temporal link prediction on realistic Ethereum data is used to evaluate the methods.Source: §I · p. 1
II. FRAMEWORK
The framework collects Ethereum transactions, constructs a Temporal Weighted Multidigraph, learns node representations with designed random-walk strategies, and evaluates them through temporal link prediction.
II. FRAMEWORK
- The framework has four procedures: data collection, network construction, random-walk-based graph representation, and link-prediction evaluation.The evaluation uses temporal link prediction on Ethereum transaction networks.Source: §II · pp. 1–2
- Collected transaction data is used to construct a Temporal Weighted Multidigraph representing Ether transfers between accounts.The records are fetched through the Etherscan API.Source: §II · p. 2
- Designed random-walk strategies learn node representations, which the framework evaluates on temporal link prediction.Source: §II · p. 2
III. DATA COLLECTION
The study collects public Ether-transfer records among external owned accounts, sampling account-centered subgraphs to limit the scale of the data. It also plans to release the collected dataset for future research.
III. DATA COLLECTION
- The dataset includes successful, nonzero-value transactions among external owned accounts (EOAs), whose transfer records are publicly available.Smart contract accounts are not included in this study.Source: §III · p. 2
- Because processing the network of more than two million EOAs is time-consuming, the study collects data around selected objective accounts as directed K-order subgraphs.K-in and K-out control the inward and outward sampling depths; Fig. 1 schematically depicts a directed K-order subgraph.Source: §III · p. 2, Figure 1 · p. 2
- The authors plan to open the collected Ethereum dataset for future studies.Source: §III · p. 2
IV. NETWORK CONSTRUCTION
The network construction models Ether transfers as timestamped, weighted directed edges, retaining multiple transactions between accounts. This representation reflects the network’s temporal evolution and transfer amounts.
IV. NETWORK CONSTRUCTION
- The construction treats transaction networks as continuously evolving and retains multiple edges because account connections represent time-dependent events.The authors also describe walks as concrete money-transfer flows and transaction amount as reflecting account similarity to some extent.Source: §IV · p. 2
- A Temporal Weighted Multidigraph (TWMDG) represents each edge as a unique directed transaction with weight and timestamp.In its edge tuple, u is the source, v the target, w the weight, and t the timestamp.Source: §IV · p. 2
- In the TWMDG, each node represents an account and each edge a unique Ether transfer; Fig. 2 depicts addresses as nodes and timestamps and amounts on edges ordered by time.Source: §IV · p. 2, Figure 2 · p. 2
V. RANDOM-WALK BASED GRAPH REPRESENTATION
The section uses random walks to represent Ethereum’s power-law transaction network while respecting temporal order and multiple transactions between accounts. It combines time- and amount-based edge-sampling strategies before learning node representations.
Random-walk motivation
- Ethereum transaction-network nodes follow a power-law degree distribution, motivating random-walk representation methods for large graphs.Fig. 3 plots the distribution for a subgraph.Source: §V · p. 2, Figure 3 · p. 3
Transaction-network representation
- Each transaction is an edge with time and amount information, and multiple transactions between accounts require preserving distinct edges rather than collapsing them into a simple graph.Temporal and weighted sampling are designed around these transaction characteristics.Source: §V · p. 3
Temporal walk construction
- Temporal walks pair nodes with traversed edges in non-decreasing timestamp order, representing feasible money-flow paths that ordinary walks may miss.Temporal successive edges provide candidate next steps for the walk.Source: §V · p. 3
Flexible sampling strategies
- The sampling design offers unbiased and biased choices in both domains: temporal bias favors edges closer to or later than the preceding transaction, while amount bias favors larger transactions.Weighted sampling uses a linear mapping to reduce the chance that small-weight edges are never sampled.Source: §V · pp. 3–4
Representation learning
- The method combines temporal and weighted selection probabilities using α, set to 0.5 by default, then learns node representations with SkipGram.The parameter balances the time and amount domains.Source: §V · p. 4
VI. LINK PREDICTION FOR EVALUATION
The evaluation uses temporal link prediction on Ethereum data, training on earlier links and testing on later links. TWMDG modeling and biased temporal walks outperform the compared approaches, with temporal walks remaining strong across varied hyperparameters.
b) Results
- TWMDG modeling outperforms traditional directed graphs, and temporal walks biased by both time and amount outperform unbiased temporal walks on temporal link prediction.Table II reports AUC and AP comparisons.Source: Table II · p. 4, §VI · p. 4
Link prediction setup
- The evaluation trains on 50% of earlier edges and tests on the remaining later edges using a support vector classifier with balanced positive and negative samples.Directed link features concatenate the source and destination node representations.Source: §VI · p. 4
c) Discussions of Results
- Temporal information and edge multiplicity are important for analyzing financial transaction networks, while time- and amount-domain information supports more comprehensive representations for prediction.Source: §VI · p. 5
d) Parameter Analysis
- Across varied settings, temporal walks outperform unbiased and biased random walks; they also remain promising across walk lengths and walks per node, while random-walk methods are more sensitive to these parameters.On EthereumG1, increasing dimension improves temporal-walk performance, whereas random-walk performance degrades above d = 64.Source: §VI · p. 5, Figure 6 · p. 5
VII. CONCLUSION AND DISCUSSION
The conclusion presents a temporal weighted multidigraph and temporal walks that account for network dynamics and edge multiplicity. Temporal link prediction experiments demonstrate the framework's effectiveness, while the learned representations cannot be obtained for newly added nodes.
Conclusion
- The proposed temporal weighted multidigraph and temporal walks incorporate network dynamics and edge multiplicity to improve graph representation.Source: §VII · p. 5
- Experiments on temporal link prediction demonstrate the effectiveness of the proposed framework.Source: §VII · p. 5
Limitations and future work
- The temporal-random-walk representation is not inductive and cannot represent newly added nodes; future work will explore other predictive tasks and larger temporal or domain-dependent networks.Source: §VII · p. 5