Multiple objective metaheuristics for feature selection based on stakeholder requirements in credit scoring
Multiple objective metaheuristics for feature selection based on stakeholder requirements in credit scoring
复制标题
DOI：
10.1016/j.dss.2021.113714
复制
发表时间：
2021-12
期刊：
影响因子：
--
通讯作者：
Naomi Simumba;Suguru Okami;A. Kodaka;N. Kohtake
中科院分区：
文献类型：
--
作者：
Naomi Simumba;Suguru Okami;A. Kodaka;N. Kohtake
替代数据越来越多地用于财务上被排斥者的信用评价。然而，当信用评估采用替代数据时，可靠性等要求变得更加重要，但这些要求并未被视为信用评分过程的一部分。本研究提出了一种在信用评分过程中纳入情境特定的利益相关者要求的方法。针对同时优化所有需求的特征选择过程，提出了两种混合启发式算法。首先是一种多目标、非支配排序、二元蚱蜢优化算法。第二种结合了遗传算法的选择、交叉和突变技术，以获得更大的多样性。两种算法均拟合由利益相关者需求得到的目标函数，进行多目标特征选择。根据利益相关者的需求和从移动、公共地理空间和卫星数据源收集的替代数据特征进行实证评估。它们的性能与几种现有算法进行了比较，它们在特定指标上提供了改进的性能。第一种算法在计算时间、收敛性和间隔方面优于现有的多目标非支配排序遗传算法NSGA-III。与此同时，第二种方法在相同的种群规模下具有更大的分布，但计算时间较长。因此，干系人的需求被成功地纳入到特征选择过程中。这样可以更好地平衡目标之间的关系。这些发现扩展了混合元启发式特征选择的研究，以及信用评分的替代数据。
Alternative data is increasingly utilized for credit evaluation of financially excluded persons. However, requirements, such as reliability, which gain new importance when alternative data is employed for credit evaluation, have not been considered as part of the credit scoring process. This research proposes an approach for incorporating context-specific stakeholder requirements in the credit scoring process. Two hybrid heuristics are proposed for a feature selection process that simultaneously optimizes all requirements. The first is a multiple objective, non-dominated sorting, binary Grasshopper Optimization Algorithm. The second incorporates the selection, crossover, and mutation techniques of genetic algorithms for greater diversity. Both algorithms are fitted with objective functions obtained from stakeholder requirements for multiple objective feature selection. Empirical evaluation is conducted with stakeholder requirements and alternative data features collected from mobile, public geospatial, and satellite data sources. Their performance is compared against several existing algorithms, and they offer improved performance on specific metrics. The first algorithm outperforms the existing many-objective non dominated sorting genetic algorithm, NSGA-III, in terms of computational time, convergence, and spacing. Meanwhile, the second method results in greater spread for the same population size but has a lengthy computational time. Thus, stakeholder requirements are successfully incorporated into the feature selection process. This results in a better balance between objectives. These findings extend the research on hybrid metaheuristics for feature selection, as well as alternative data for credit scoring.