Abstract: In the context of online lending, Adaboost based credit assessment approach has high disagreement level among different basic classifiers and high misclassification cost on sample set. Existing studies fail to consider the effect of basic classifiers’ disagreement and samples’ misclassification cost on the sample weights in basic classifiers, leading to the decline of the performance of credit assessment model in online lending businesses. Therefore, we propose a credit assessment approach based on an improved Adaboost model. This approach can improve the strategy of weighting in Adaboost model through considering the misclassification rates of basic classifers, the disagreement of learning results of samples among basic classifiers, and the misclassification cost of samples. Our approach can improve the validity of credit assessment results in online lending context by focusing on the samples that are difficult to classify and have high misclassification cost. The experimental results on PPDai platform show that the approach has higher accuracy and lower misclassfication cost than traditional Adaboost based credit assessment model.

Key words: credit assessment approach, adaboost, disagreement, misclassification cost

摘要： 网络借贷环境下基于Adaboost的信用评价方法具有较高的基分类器分歧度和样本误分代价。现有研究没有考虑分歧度和误分代价对基分类器样本权重的影响,从而降低了网络借贷信用评价结果的有效性。为此,提出一种基于改进Adaboost的信用评价方法。该方法根据基分类器的误分率,样本在不同基分类器上分类结果的分歧程度,以及样本的误分代价等因素,调整Adaboost模型的样本赋权策略,使得改进后的Adaboost模型能够对分类困难样本和误分代价高的样本实施有针对性的学习,从而提高网络借贷信用评价结果的有效性。基于拍拍贷平台数据的实验结果表明,提出的方法在分类精度和误分代价等方面显著优于传统的基于Adaboost的信用评价方法。

关键词: 信用评价方法, Adaboost, 分歧度, 误分代价