Addressing the cold-start problem of recommendation systems for financial products by using few-shot deep learning

With advances in machine learning, the application of financial services has become increasingly intelligent. The artificial intelligence financial advisory and other customized services have been launched in the market. However, when facing new financial products, recommendation models may not possess sufficient training data, which is called the cold-start problem. To address this problem, we present the few-shot learning model, which aims to learn a new category through a limited number of samples by transferring the past experiences and knowledge to the learning of new categories. We trained our model by using considerable past transaction data and applied a customized bias to adjust the predicted result for each investor. We then tested our model on new stock items by applying few-shot learning and testing. The results of this study indicate that training the proposed network with a customized bias can improve its prediction accuracy, especially for new items. Moreover, the proposed model can transform past experience into knowledge. When new data with only a few available samples are input into the proposed model, it can learn and make predictions and thus achieve few-shot learning.