金融大数据迁移学习及其应用

Transfer Learning and Its Applications in Financial Big Data

在金融科技发展的背景下,本文从金融大数据复杂特征入手,阐述了使用多源数据信息辅助目标任务的重要作用,从多源数据的角度解释了迁移学习技术对处理数据异质性的重要性,并对迁移学习技术相关概念和方法进行梳理和综述,包括数据驱动和基于模型的迁移学习方法.此外,本文提出了基于广义矩估计(GMM)的迁移学习方法的统一框架,给出有效的迁移学习的算法,并将所提出方法应用于多源数据下基于风险价值(VaR)和期望分位数(expectile)的风险度量.然后,在银行个人信用评估场景下模拟了样本量不足和样本不平衡两种情况,测试了三种迁移学习方法的预测精度,并分析了筛选资源域信息对提高估计精度的重要作用.最后,描述了更多迁移学习在金融领域的应用场景与发展前景.

In the context of the development of financial technology,we start with the complex characteristics of financial big data and elaborate on the importance of transfer learning of using multi-source data information to assist target tasks.We explain the significance of transfer learning technology in dealing with data heterogeneity from the perspective of multi-source data,and summarize the relevant concepts and methods of transfer learning technology,including data-driven and model-based transfer learning methods.In addition,this paper proposes the unified framework of transfer learning method based on generalized moment estimation(GMM),gives the effective algorithm of transfer learning,and applies the proposed method to the application of transfer learning in risk value(VaR)and risk measure based on expected quantile(expectile)under multi-source data.Then,we simulate two scenarios where samples are of insufficient or imbalanced sample sizes,respectively,in the application to personal bank credit evaluation,with tests of the prediction accuracy of three transfer learning methods,and analysis of the importance of filtering resource domain information.Finally,we described more application scenarios and development prospects of transfer learning in the financial field.