A local binary social spider algorithm for feature selection in credit scoring model

Abstract

Borrowers’ default is the main risk of online lending. Using credit scoring model to assess borrowers’ credit is an important means to reduce default risk. The existing related work focus on the improvement of assessment methods, and do not study the quality of credit data enough. Actually, there are usually some noisy, redundant or irrelevant features in online credit data, which increase the computational complexity and reduce the evaluation accuracy of the model. Well-performing feature selection method is the key premise to improve accuracy of the credit evaluation model. At present, the feature selection methods applied to online credit scoring generally have shortcomings such as subjectivity, time consuming, low accuracy, etc. So it is urgent to introduce new schemes. Among the feature selection schemes in many fields, heuristic algorithm is widely recognized because of its advantages of high efficiency and accuracy, and especially BinSSA is an outstanding representative. However, BinSSA has the defects that it is easy to form extreme distribution in initialization, and is inefficient because there is a high probability of falling into local optimum during iteration. So it is still not the best choice for feature selection in online lending credit scoring model. In this paper, we propose a local binary social spider algorithm (LBSA), which introduces two local optimization strategies into BinSSA: opposition-based learning (OBL) and improved local search algorithm (ILSA). These strategies can help to improve the above defects. We conduct comparative experiments based on three typical online credit data sets and different algorithms, to verify the superior performance of LBSA. The experimental results validate that LBSA greatly reduces the redundancy of returned feature subsets, improves the iterative stability, and makes the credit scoring model more accurate and effective.