基于多源多任务深度学习实现个性化金融推荐的算法研究

The research of financial recommendation systems aims to recommend financial products or portfolios to individuals and micro-, small and medium-sized enterprises. It makes a lot of sense for inclusive finance. However, previous studies always only use single-source data and rarely consider users' personalized investment preferences. These limitations will limit the performance and practicability of financial recommendation systems..In response to the above problems, this project studies algorithms of financial recommendation system which based on deep-learning, multi-source learning and multi-task learning theories. The algorithms will use multi-source data for learning a financial recommendation system which can recommend high-quality and personalized financial products to users..First, this project will propose methodologies which utilize financial data from multiple sources and users' personalized data to learn a financial recommendation system. Then, this project will study how to discover high-quality financial products which also match users' personalized investment preferences. Moreover, this project will also propose strategies to avoid financial recommendation systems over-fitting by considering that financial data are always time-varing and have a low signal-to-noise ratio. Finally, the financial-recommendation algorithms proposed by this project can combine multiple sources of data and recommend financial products which not only are with high-qualities but also match users’ personalized preferences. Moreover, this project will also develop real-world applications corresponding to the algorithms of financial recommendation systems. Thus, this project provides highly important academic and economic values.

金融推荐算法能实现自动向中小型投资者推荐金融产品和理财建议，这对推动普惠金融的发展有着重要意义。然而，传统金融推荐算法通常基于单一数据源且很少考虑用户的个性化投资需求。上述缺点大大限制了金融推荐系统的效果和实用性。.本项目基于深度学习理论开展金融推荐算法的研究。同时结合多源学习和多任务学习两个理论，研究基于多源异构数据源并且同时考虑到金融产品质量和投资者个性化投资需求的金融推荐算法。.本项目将在深度神经网络的输入端根据各类金融数据源以及用户数据的类型和特点研究编码层和融合层，并研究在输出端能够同时考虑金融产品质量和用户个性化投资需求的多任务学习算法。本项目也将在深度神经网络训练方式上考虑金融数据时变性和低信噪比的特点，设计相应的避免过拟合的策略。最终，本项目能实现适用于大数据时代以及个性化时代的金融推荐算法。另外，项目也将开发相应的原型系统。因此，本项目的研究成果具有重要的学术和经济价值。

作为保护中小投资者利益的重要工具，金融推荐系统的研究具有至关重要的应用价值。然而传统的方法多数无法应对金融领域数据多源异构的特点，且无法很好地实现个性化金融产品配置的需求。本课题基于机器学习理论开展研究，实现能够同时从高质量和个性化两个目标任务学习模型的金融推荐算法。项目的研究属于金融推荐系统、深度学习、多源学习与多任务学习的交叉话题，实现从上述三个方向突破传统金融推荐算法的局限。课题组发表标注本项目的论文8篇，指导博士生若干。并且本项目也基于Python搭建了金融推荐系统。这些研究成果成功地将金融产品推荐系统的问题刻画成为是一个典型的机器学习问题，并且在特征工程方面做出创新，大幅度提升了金融产品推荐系统的性能。