LLP-AAE 算法在金融风险识别领域的应用-《计算机技术与发展》

文章信息/Info

Title:: Adversarial Auto-encoding for Learning from Label Proportions on Financial Risk Identification

Author(s):: FAN Zi-yi; YANG Huan; School of Information,University of International Business and Economics,Beijing 100029,China

Keywords:: AAE; LLP; financial risk identification; weak supervised learning

摘要:: 通过结合机器学习中的标签比例学习问题与生成模型算法,提出了一种新的深度学习算法,称为对抗性自动编码器算法。考虑到金融领域的数据具有天然的获取成本高、保密性大与客户可分类的特点,故选择把对抗性自动编码器算法实践于银行的风险检测中。在标签比例学习中,数据集为包层面的标签比例数据,将这种类型的数据作为输入,对抗性自动编码器依赖编码器解码器等的对抗训练来捕获原始数据分布,并得到足够逼真的合理重构分布。同时施加一种对抗学习机制作为一种正则化,该部分对抗使得编码器学习足够充分与多维,进一步提升了算法的效果。与现存的方法相比,提出了具有可扩展性的深度学习解码器,拓宽了应用范围情景,侧重将标签比例学习领域的深度学习方法应用于金融类型数据之中。

Abstract:: Combining a weakly - supervised method,learning from label proportions ( LLP) and a generative model,adversarial auto -encoders ( AAE) ,we propose an effective deep learning algorithm of tabular data, called LLP via adversarial auto - encoders ( AAE -LLP) . Considering that the data in the financial field are naturally characterized by high acquisition cost, high confidentiality and customer classification,the adversarial auto encoder algorithm is chosen to be applied in the risk detection of banks. In the tag proportion learning,the data set is the tag proportion data at the packet level,and this type of data is taken as the input. Adjective auto encoders rely on the antagonism training of encoders and decoders to capture the original data distribution, and get a reasonable reconstructiond istribution that is realistic enough. At the same time,an antagonistic learning mechanism is applied as a kind of regularization,which makes the encoder learning sufficient and multi - dimensional, and further improves the effect of the algorithm. Compared with the existing methods,a deep learning decoder with extensibility is proposed to expand the scope of application scenarios, and the deep learning method in the field of label proportional learning is applied to financial type data.