Research And Implementation Of P2P Lending Agency Risk Evaluation Techology Based On Reinforcement Learning

In recent years,P2P network lending has developed rapidly in China and the scale has become the first in the world.P2P network lending has brought convenience to both the public and investors.With the rapid development of P2P lending agencies,the risks of these agencies have gradually emerged.The risks mainly include three forms:out of business,difficult withdrawing and absconded with ill-obtained gain.The risks of these P2P network lending agencies often have a negative impact on society,hence it is necessary to evaluate the risks automatically.Although machine learning can be used to automatically evaluate the risks of P2P agencies,most of machine learning methods require high data quality.Currently,the data scale of P2P agencies is small and the data distribution of various risk types is imbalance.Hence,in order to address the issues of small-scale data and data imbalance of P2P agencies for text categorization,this paper introduces a reinforcement learning model.Through the research and analysis of various reinforcement learning algorithms,this paper proposes a P2P network lending agency risk evaluation approach based on reinforcement learning,which considers the risk evaluation task as a kind of text classification.However,the existing reinforcement learning model is so time-consuming that it is unable to gain realistic applications.Therefore,this thesis proposes an improved reinforcement learning model enlightened by the idea of experience replay and priority sampling,which dynamically update the sample weights so as to speed up the model training substantially.Based on the above technology,this paper proposes a research framework for P2P lending risk evaluation approach based on reinforcement learning,including preprocessing,text feature representation,reinforcement learning baseline model construction,risk evaluation and model training improvement.Afterwards,the validity of the model is verified by experiments.Experimental results show that the reinforcement learning model proposed in this thesis is less affected by data imbalance than the traditional machine learning model and can achieve better results.Finally,this thesis implements an experimental system for P2P lending agency risk evaluation based on the proposed reinforcement learning model.