Adaptive Multi-objective Optimization-based Ensemble Classification Method And Its Application In Credit Scoring

With the development of artificial intelligence technology,academia has paid more and more attention to all aspects of machine learning research,and various industries have begun to use machine learning technology to empower businesses.The traditional machine learning methods are usually based on the assumption of balanced data distribution.In actual application scenarios,the problem of imbalanced data distribution often occurs,resulting in a certain impact on the performance of traditional machine learning methods.In response to this problem,the topic of imbalance learning has been proposed and satisfactory performance has been achieved on imbalanced datasets.Credit scoring is one of the important application scenarios of imbalance learning.In the credit scoring business,the machine learning model is used to predict the customer's credit qualification.Its advantage is that it can quickly make predictions on the customer's credit to assist business decision-making.It plays a large role in constructing customer portraits,risk control,and marketing business.However,due to the complexity of credit scoring business and the existence of heterogeneous data,the problem of data imbalance in the context of credit scoring has become a more challenging research topic.In recent years,imbalance machine learning methods in credit scoring scenarios mainly include sampling methods,cost-sensitive methods,imbalance ensemble methods,etc.These methods have achieved good performance in credit scoring scenarios,but there are still corresponding limitations,mainly manifested in: 1.Sampling methods usually lack of the optimization of sample distribution,and sparse sampling-based methods usually lose some useful information;2.Some imbalance methods such as cost-sensitive methods are sensitive to noise samples and outliers,and the performance is not stable;3.The imbalance ensemble classification methods usually lack of adaptive optimization mechanism,and few of them consider the combined optimization of samples and feature spaces,etc.Based on the limitations mentioned above of imbalance learning,this paper proposes an adaptive multi-objective optimization ensemble classification method for credit scoring.The innovations of this method include: 1.A particle swarm optimization algorithm is introduced into the existing oversampling method to further optimize the sample distribution after sampling,so that the sample distribution after oversampling is closer to the original data distribution;2.Introduce both sample dimension optimization and feature dimension optimization into the ensemble learning framework.After performing sample optimization,multi-objective optimization algorithm combined with common evaluation indicators of credit scoring scenarios is used to optimize the random feature subspace and perform ensemble classification to improve model performance;3.Design an adaptive optimization strategy to optimize the classifier set,and draw on the idea of reinforcement learning feedback mechanism to update the probability of the classifier being sampled in the corresponding feature subspace,and finally get a better classifier set.In the experiment,detailed comparative experiments on the key technologies of the algorithm are made,and their optimization effects through data visualization are also displayed.Experimental results on real-world credit scoring datasets show that the overall performance of the proposed algorithm is better than the current mainstream imbalance classification algorithm.