Research On Stock Recommendation Method Based On Relation Network

With the rapid development of the financial market,the stock market has become the focus of investors,and investors trade in the stock market to obtain the investment returns.The stock recommendation task is to recommend stocks with higher return ratios for the investors to reduce investment risks.In the past,most stock prediction methods study the historical sequence patterns or financial text features to predict stock trend in the near future.Some studies have begun to focus on exploiting the relation between stocks for stock prediction and recommendation recently.However,due to the time-varying relation strengths between stocks and the different relation types in the stock relation network,the existing models cannot accurately obtain the relation information between stocks,which limits the accuracy of stock prediction.Therefore,this thesis takes the time-varying and the different types of stock relation data into full consideration,and designs the stock recommendation models based on graph neural network for two stock relation networks.The main work and contributions of this thesis are as follows:(1)Considering the dynamic time-varying strength of the industry relation among stocks,this thesis proposes a time-aware relational attention network for graph-based stock recommendation according to return ratio ranking.This model can capture the time-aware relation strengths from the dynamic interaction of stock historical sequence information and stock description document information,and use a graph convolution algorithm to aggregate the features of adjacent stock nodes with time-aware relation strengths.Finally,according to the historical features and relational features of the stocks,the expected return ratios of all stocks are predicted,and the stock with the highest expected return is recommended to investors.Experimental results on several American stock market datasets show that the proposed model outperforms other stock recommendation methods.(2)Considering that multiple types of stock relations in the stock knowledge graph will have different effects on stock prediction and recommendation,this thesis proposes a stock recommendation model based on a relation-aware attention graph convolutional network.This model captures deeper relation features in the stock knowledge graph by capturing the attention coefficients between neighbour stock nodes and focusing on different types of stock relations for information aggregation.Finally,the stock recommendation is carried out by combining the historical features of the stock historical sequences.This model takes full advantage of the multi-type relational information of the stock knowledge graph and outperforms other methods in stock recommendation experiments in American stock market datasets.