Abstarct: Identifying the different behavior and pattern of data among the large amount of information available is of great importance. Nowadays, one of the most important areas for identifying anomalies in data is the financial field and conducting transactions and sales. Due to the high volume of transactions, the large number of investors and the complexity of the transaction process, fraud and economic collusion are among the problems that exist in this sector. With the establishment and growth of e-government in many countries, many tenders and auctions conducted in the public sector go beyond the traditional methods and are conducted in the context of information technology and electronically. The present study was conducted with the aim of clustering and identifying suspicious individuals in public sector tenders and has investigated the effect of using deep learning methods representing the trading graph in identifying suspicious data behavior. In this study, with the help of clustering method, customer behavior in transactions was analyzed and data pattern in different clusters was analyzed. The set of data used in the present study includes tenders conducted by government agencies in Iran, which is available to the public in the period of 1389 to 1398 in the government procurement system. After modeling the data in the form of trading graphs, important graph indicators were extracted by network analysis methods and valuable trading graph associations were extracted and suppliers and executive agencies influencing large transactions were identified. Graph nodes were embedded using graph representation deep learning approaches to implement clustering algorithm and cluster analysis on associations data. The characteristics of each trading node (amount of value of the transaction and the number of transactions) along with the characteristics obtained through graph analysis and graph embedding feature vector, were considered as data input of the clustering algorithm. Finally, by analyzing the clusters and data within each of those clusters, suspicious data with different behavior were identified. The research findings show that the use of graph properties along with the embedded feature vector of graph nodes obtained by deep learning methods of graph representation, contributed significantly to better and higher quality clustering of data. In addition, this study showed that the longer the feature vector in graph embedding, the better the data clustering performance due to the availability of more information of neighboring nodes in the graph, and the more accurate the data is identified with suspicious behavior among the data. .