Research On Knowledge Graph Representation Learning And Its Application In Stock Price Prediction

Knowledge graph representation learning is widely used in all walks of life by mapping entities and relationships in the graph to low-dimensional continuous vector space,thus preserving structural and semantic information while facilitating computation,and better serving downstream tasks,such as intelligent question answering and recommendation systems.Moreover,due to the abundance and complexity of financial data,it contains a large amount of discrete information and the interrelationships are not clear.Therefore,knowledge graph representation learning techniques can be applied to financial data related applications to better mine available information in financial data.Based on this,the paper uses stock data as an example to study the construction and application of knowledge graph of financial data.The main tasks of this paper include applying knowledge graph to the task scenario of stock price prediction,so as to integrate the multi-factor characteristics affecting stock prices and improve the performance of model.Furthermore,in response to the contribution of knowledge graphs in improving the performance of stock price prediction,this paper conducts a more in-depth study on knowledge graph representation learning to improve the representation ability of data information.So it can be better used in financial data mining and analysis.Specific research contents are as follows:Firstly,how knowledge graph can provide more comprehensive knowledge information for stock prediction is studied.Due to the numerous influencing factors in the stock market,there are complex nonlinear interactions between different factors.According to the research,the prediction of a single stock usually ignores the influence of relevant market information.Therefore,in order to improve the accuracy of stock price prediction,this paper introduces knowledge graph and graph embedding technology,based on which market information can be obtained to correlate stocks,as one of the important influencing factors.At the same time,since the stock price is not stable and many factors will cause the stock price mutation,resulting in the general prediction mis-accuracy,this paper proposes the mutation point detection subnetwork,through the construction of mutation point distance weight matrix,to obtain the stock price mutation characteristics,so that the model will not cause learning errors due to the existence of mutation points.Considering that the existence of mutation points will cause a large deviation between the real value and the predicted value,this paper introduces a piecewise loss function to eliminate large errors.To sum up,the new prediction scheme covers relevant factors as far as possible,and improves the performance of stock price prediction by integrating the three key features of target stock price characteristics,stock market characteristics and mutation point characteristics.Secondly,the knowledge graph representation learning model is studied.In general application,knowledge graph still has the problem of insufficient ability to extract higherorder feature representation,and the existence of sparsity in its data also has an impact on the accuracy.Therefore,in order to mine the Semantic information and structural information in the knowledge graph more accurately and deeply,this paper uses the selfsupervised comparative learning mechanism to extract features from the three kinds of views:knowledge graph,hypergraph and line graph,so as to improve the knowledge graph representation learning model.In order to verify the validity of the model proposed in this paper,we continue to apply it to the task of stock price prediction,and improve the stock price prediction model by integrating the local,semi-global and global features of the extracted stock information.It was shown through experiments that the representation learning model proposed in this paper improves the representation ability of stock related features in the downstream task of stock price prediction,and its performance is better than many previous methods.