以太坊是一个开源的、支持智能合约的区块链平台，截止至今已经积累了数十亿条交易记录，以太坊上所有的交易数据都是开放的，这给研究人员一个很好的机会去分析以太坊中的数据。以太坊从建立到现在，已经有很多关于安全、性能等方面的研究，但是关于挖掘以太坊中用户间关系的研究还不够成熟，用户间的行为包括创建和调用智能合约，互相转账，如何利用用户间现有的行为数据进行建模，预测异常用户未来的行为，有利于人们对以太坊有更深的了解。另外，2022年3月最高人民法院明确把虚拟币交易列为非法吸收资金的行为方式，因此如何利用现有以太坊中的数据预测未来可能发生的交易，能够帮助人们追踪以太币的实时交易轨迹，提前预警非法用户的危险行为。为了更好地分析以太坊交易网络，本文获取了以太坊中（2015-07-30～2021-04-01）所有上千万个区块的内容和其中的交易数据，把整个以太坊历史交易细分为若干个时间窗口，从网络演化的角度详细观察和分析了以太坊交易网络，并根据以太坊交易网络具有丰富的时间和空间信息的特点，提出了两个适合以太坊交易网络的建模方法，分别是基于时空随机游走的以太坊交易网络建模和基于交易动态图嵌入的以太坊交易网络建模。第一个模型将以太坊交易网络抽象成了一个基于时间的有向多边图，更全面地表示账户之间的交易转移行为，充分利用了以太坊具有丰富时间信息的特性，并在下游任务，以太坊未来交易预测中进行了实验，实验表明，通过综合考虑时间信息与空间信息的建模方法的效果明显优于先前只考虑空间信息的研究。第二个模型将以太坊交易网络抽象成了一个基于交易的连续时间动态图，并使用基于交易动态图嵌入的方法获得每个账户的动态嵌入表示，解决了前一个模型只能应用在直推式学习而带来的无法适用于不断演化的大规模图场景等问题，并针对以前方法聚合邻居信息性能较差的缺点，提出了High-Order Attention有效聚合了高阶邻居节点的信息，本文同样在以太坊未来交易预测任务中进行了实验，实验表明，这种建模方式适合归纳式学习的方式，可以用于预测以太坊未来某具体时间点的交易，且预测性能也优于本文先前提出的模型和一些其他动态图模型。

Ethereum is an open-source blockchain platform that supports smart contracts. So far, billions of transaction records have been accumulated. All transaction data on Ethereum is open, which gives researchers a good opportunity to analyze data in Ethereum.From the establishment of Ethereum to the present, there have been many researches on security and performance, but the research on mining the relationship between users in Ethereum is not mature enough. Users are creating and calling smart contracts, and users are also transferring money to each other. How to use the existing behavior data among users to model and predict the future behavior of abnormal users will help people to have a deeper understanding of Ethereum. At the same time, Moreover, in March 2022, the Supreme People's Court clearly listed virtual currency transactions as a way of illegally absorbing funds.Therefore, how to use existing data in Ethereum to predict possible future transactions, which can help people track the real-time value of Ethereum Transaction track, early warning of dangerous behavior of illegal users. In order to better analyze the Ethereum transaction network, this thesis obtains the content and transaction data of all tens of millions of blocks in Ethereum (2015-07-30$\sim$2021-04-01). The historical transaction is subdivided into several time windows, and the Ethereum transaction network is observed and analyzed in detail from the perspective of network evolution. The modeling methods are the Ethereum transaction network modeling based on temporal-weighted random walk and the Ethereum transaction network modeling based on transaction dynamic graph embedding.The first model abstracts the Ethereum transaction network into a Temporal Multidigraph, which more comprehensively represents the transaction transfer behavior between accounts, and makes full use of the rich time information of Ethereum. Experiments have been performed in the downstream task of Ethereum transaction prediction. The experiments show that the preform of the modeling method by comprehensively considering temporal information and spatial information is significantly better than the previous research that only considers spatial information.The second model abstracts the Ethereum transaction network into a continuous-time dynamic graph, and uses the method TGE to obtain the dynamic embedding of each account, which solves the problem that the previous model can only be applied to transductive learning and not suitable for the continuously evolving large-scale graph. In view of the disadvantage of poor performance of the previous methods for aggregating neighbor information, High-Order Attention is proposed to effectively aggregate the information of high-order neighbor nodes.We also do experiments in the task of Ethereum transaction prediction using this model. The experiments show that this modeling method is suitable for the inductive learning, which can be used to predict the transactions of Ethereum at a specific time point, and the perform is also better than the model proposed earlier in this thesis and some other dynamic graph models.