Financial news recommendation based on graph embeddings

Highlights：

• We introduce NNR, a graph-based method for financial news recommendation that outperforms the commonly used baselines.

• We develop INNR, an incremental approach that makes the recommendation process more efficient.

• We combine NNR and INNR to create a graph embeddings-based financial news recommendation framework that can achieve a good balance between time efficiency and recommendation accuracy in the financial news recommendation task.

摘要

Most of the existing methods are not enough for securities companies to make their decisions on recommending the most suitable financial news to a specific user. On the one hand, such news articles often contain external knowledge related to companies and stocks. On the other hand, it is important for financial news recommendations to dynamically measure users' interests since people are usually interested in multiple specific concepts, companies, stocks and industry categories. To address the above challenges, we start by building a heterogeneous graph consisting of users, news, companies, concepts, and industry categories. Then, the graph embeddings of the nodes are generated using node2vec, and user-news relatedness can be computed based on them. Since financial news articles are time-sensitive, we propose an incremental method for alleviating the computational efficiency problem. The combination of a node2vec-based recommendation method and the incremental method can achieve a good balance between time efficiency and recommendation accuracy in the financial news recommendation task. Our methods are evaluated on a real-world dataset from a Chinese securities company, are shown to outperform other commonly used baseline models, can provide decision support for companies choosing news to be recommended to target users and allow users to obtain personalized real-time news recommendations.