Application And Research On Network Representation Learning In Financial Equity Network

In the last 30 years,the financial markets in China have made remarkable achievements.Financial institutions have accumulated the mass transactions and information data with the development of Internet technology and the continuous improvement of service quality.And how to make comprehensive use of these data in order to better feedback to the financial markets is an urgent problem to solve for financial enterprises.In the traditional economic field,various economic indicators are usually used to study and analyze the market behavior and market appearance with the great progress of complex system theory making the research of complex network obtain a lot of rich results.Therefore,it is of great theoretical and practical significance to apply the research results of complex network to financial markets,especially under the background of China's vigorous promotion of fintech strategy,the research of financial data through complex network analysis of complex networks has become a hot topic in the industry and academia.The financial equity network makes up for the lack of equity investment information in the traditional financial network research,as an important part of the financial markets,and more objectively depicts the complex characteristics of the financial markets.The structure of stock ownership is very complicated and includes stock,trust,fund,futures and other financial derivatives.Traditional complex network technology usually uses custom network characteristics for analysis.However,such kind of manually extracted data feature is coarsegrained,which is difficult to mine the deep value of the network.Deep learning has been a huge success in the industry in recent years,and people pay more and more attention to end-to-end learning and representation learning.Network representation learning technology vectorizes complex graph-structured data,while graph convolutional neural network realizes the end-toend learning of graphs,which greatly enriches the technical system of graph-based data mining.Therefore,the use of network representation learning technology can further obtain the deep value of financial equity network.The financial equity network structure is complex and changeable,we introduce the concept of capital community to divide the financial equity network to further reveal the internal structure of the network,which can provide new ideas for stock classification and equity relationship description.Based on the large amount of shareholding data disclosed by listed companies,this thesis discovers the capital community in the financial equity network,and reveals the relationship between financial derivatives in the financial network community.After that,the effect of algorithm division is further improved with the help of the fusion of investment transaction information of nodes,and end-to-end learning training is realized for other financial networks.The main work of this thesis is as follows:(1)This thesis proposes a random-walk strategy with network topology coding in view of the special nature of the topological structure of financial equity network,and considers that the classical random-walk sampling strategy is difficult to reflect the differences between node structures.And the hash mapping function is used to optimize the sampling process of nodes,so as to improve the sampling efficiency of nodes,and make the embedded vectors of nodes continue to save the topological structure information.(2)In this thesis,a random-walk based node embedding algorithm is designed and implemented in dynamic network.In view of the temporal characteristics of the financial equity network,the edge weights and node attributes of the network are integrated into the sampling process,so that the represented vectors can better capture the topological structure properties of the temporal network.Also this thesis extends the algorithm and explores the value of network datasets under different backgrounds.(3)Aiming at the large amount of structured and unstructured data information in the financial equity network,this thesis uses GraphSage algorithm to fuse the quantified information of nodes into the vector representation.Also we design a mini batch temporal node aggregation and expansion mode for large scale dynamic network datasets.According to GraphSage we extend graph convolutional neural network algorithm to the dynamic network.Then we design the unsupervised loss function training algorithm model for the specific financial equity network mining and finally get the obtained experimental results compared and analyzed.