Edelman, J. Baesens, D. Roesch, H. Skip to content. Star Branches Tags. Could not load branches. Could not load tags. Latest commit. Git stats 72 commits. Failed to load latest commit information. Two-level classifier ensembles for credit risk assessment. Expert Systems with Applications, 39 12 , — Tripathi, D. Hybrid credit scoring model using neighborhood rough set and multi-layer ensemble classification.
A comparative study on base classifiers in ensemble methods for credit scoring. Expert Systems with Applications, 73 , 1— Parvin, H. Proposing a classifier ensemble framework based on classifier selection and decision tree. Engineering Applications of Artificial Intelligence, 37 , 34— Saha, M. Credit cards issued. Vapnik, V. The nature of statistical learning theory. NY: Springer. Cortes, C. Support-vector networks. Machine learning, 20 3 , — Van Gestel, T. Bayesian kernel based classification for financial distress detection.
European journal of operational research, 3 , — Yang, Y. Adaptive credit scoring with kernel learning methods. European Journal of Operational Research, 3 , — Zhou, L. Credit scoring models with auc maximization based on weighted svm.
XIAO, W. A study of personal credit scoring models on support vector machine with optimal choice of kernel function parameters [j]. Li, S. The evaluation of consumer loans using support vector machines. Expert Systems with Applications, 30 4 , — West, D. Neural network credit scoring models. Haykin, S. Neural networks: A comprehensive foundation. NY: Tsinghua University Press. Atiya, A. Bankruptcy prediction for credit risk using neural networks: A survey and new results.
IEEE Transactions on neural networks, 12 4 , — Evolutionary extreme learning machine with novel activation function for credit scoring. Engineering Applications of Artificial Intelligence, 96 , Binary bat algorithm and rbfn based hybrid credit scoring model. Multimedia Tools and Applications, 79 43 , — Bat algorithm based feature selection: Application in credit scoring. A new hybrid ensemble credit scoring model based on classifiers consensus system approach.
Expert Systems with Applications, 64, 36— Yeh, I. The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients. Expert Systems with Applications, 36 2 , — A comparative assessment of ensemble learning for credit scoring.
Expert systems with applications, 38 1 , — Nanni, L. An experimental comparison of ensemble of classifiers for bankruptcy prediction and credit scoring. Expert systems with applications, 36 2 , — Zhang, D.
Vertical bagging decision trees model for credit scoring. Expert Systems with Applications, 37 12 , — Lin, W. Machine learning in financial crisis prediction: a survey. Lahsasna, A. Credit scoring models using soft computing methods: A survey. Abdou, H. Credit scoring, statistical techniques and evaluation criteria: a review of the literature.
Intelligent Systems in Accounting, Finance and Management, 18 2—3 , 59— Extreme learning machines for credit scoring: An empirical evaluation. Expert Systems with Applications, 86 42— Classifiers consensus system approach for credit scoring. Knowledge-Based Systems, , 89— Tsai, C.
Using neural network ensembles for bankruptcy prediction and credit scoring. Expert systems with applications, 34 4 , — Xia, Y.
A novel heterogeneous ensemble credit scoring model based on bstacking approach. Expert Systems with Applications, 93 , — Guo, S. A multi-stage self-adaptive classifier ensemble model with application in credit scoring. IEEE Access, 7 , — Wongchinsri, P. Sr-based binary classification in credit scoring , — IEEE. Hens, A.
Computational time reduction for credit scoring: An integrated approach based on support vector machine and stratified sampling method. Expert Systems with Applications, 39 8 , — Huang, C.
A ga-based feature selection and parameters optimizationfor support vector machines. Expert Systems with applications, 31 2 , — Hu, Q. Neighborhood rough set based heterogeneous feature subset selection. Information sciences, 18 , — Liu, Y. An improved particle swarm optimization for feature selection.
In order to analyze the impact of building fair scoring models the data have been extended by an artificial protected variable Gender to mimic A9. For this purpose the variable Status with largest IV has been selected to construct the new protected variable.
As it is shown in Table 3 , in the first step two of the status-levels are assigned to women and the other two variables are assigned to men, respectively. In consequence, women take the lower risk compared to men from their corresponding status levels in the artificial data. The resulting graph is. Information values of the scorecard variables and the removed variable personal status and sex.
As Lemma 1 of Kusner et al. The idea can be generalized for larger sets of X , P. As a result, the designed effect of gender on status is disturbed to some extent and only holds for the remaining observations. Logistic regression still represents the gold standard for credit risk scorecard modeling Crook et al.
For this reason, a traditional scorecard using logistic regression is created as a baseline model for the simulation study. For plausibility reasons the bins of four of the variables Duration, Employment duration, Amount, as well as Age are manually updated where only one of them Duration did enter the final model after BIC based stepward forward variable selection cf.
Figure 1. Automatically created bins A for the variable duration and manual update B : a plausible trend of increasing risk with increading duration. For plausibility reasons i. Table 4. After forward variable selection using BIC on the training data the resulting scorecard model uses five input variables as they are listed in Table 2.
The equation of the resulting logistic regression model is given in Table 5. The corresponding scorecard model with frequencies and default rates for all classes can be found in Table 6. TABLE 6. Resulting scorecard model. In practice it is usual to assign scorecard points to the posterior probabilities as given by the score. In addition to the traditional scorecard baseline model fair models are developed according to the algorithm presented in Section 2.
Note that for companies depending on the business strategy and corresponding acceptance rates it can be more suitable to put more emphasis on other performance measures such as the partial AUC Robin et al. In order to compute the GUI a cut off s 0 for the score S has to be defined: For the scope of the simulation study within this study the cut off has been set to the portfolio default rate 0.
In practice, rejecting all customers with a risk above average will lead to an unrealistically high rejection rate. Both results are given in Table 7. Fairness-performance trade-off: performance solid and unfairness dashed for traditional blue and fairness-corrected green model for different levels of correlation between the protected variable gender and the prediction variable status. The red dotted line indicates the thumb rule for unfairness.
TABLE 7. The solid lines indicate performance on the test data for the traditional blue and the fair model green. The traditional model is unaffected by the protected variable and thus of constant performance with a Gini coefficient of 0. The corresponding dashed lines show the group unfairness index of both models where the additional dotted red line represents the thumb rule threshold of 0.
Figure 3 shows an example of the partial dependence profiles cf. Section 3. Table 7. For these data no strong differences in performance are observed 0. Protected attribute dependence plot of the traditional scorecard A vs. Along with these promising results another side effect can be noticed: As a consequence of gender-wise fairness correction there are different WOEs for both genders in all bins and consequently also different scorecard points for each gender as it can be seen in Table 8.
Thus, the price of having a fair scoring model is different points with respect to the protected attributes here: Gender.
Not enough, a traditional plausibility check during the scorecard modeling process concerns monotonicity of the WOEs with respect to the default rates Szepannek, which now has to be done for all levels of the protected attribute and is not necessarily given anymore after fairness correction.
Note that also in our example the order of the default rates of the two bins with the highest risk of the variable Status has changed for the female customers. TABLE 8. Comparison of the variable Status for the traditional model left and female center and male right gender in the fair model.
Although in general the presented methodology can be applied to arbitrary machine learning models the changes in the data as induced by the fairness correction put even more emphasis on a deep understanding of the resulting model and corresponding methodology of interpretable machine learning to achieve this goal cf. Further note that as it is demonstrated in Szepannek the obtained interpretations bear the risk to be misleading.
For this reason other authors such as Rudin suggest restricting interpretable models and in summary a proper analysis of the benefits of using more complex models should be done in any specific situation Szepannek, Then all descendants of the protected attributes must be corrected accordingly. In this study, different definitions of fairness are presented from the credit risk scoring point of view as well as a fairness correction algorithm based on the concept of counterfactual fairness. Furthermore, the idea of population stability is transferred into a new group unfairness index which allows quantifying and comparing the degree of group fairness of different scoring models.
In addition, partial dependence plots are proposed to visualize the fairness of a model with respect to some protected attribute. Based on these measures, a simulation study has been set up which makes use of a corrected version of the well-known German credit data. The results of the study are quite promising: Up to some degree fairness corrections are possible without strong loss in predictive accuracy as measured by the Gini coefficient on independent test data.
The explanation of algorithmic decisions gets even more complicated and future work has to be done in order to investigate the observed effects of our study for other classes of machine learning models such as random forests, gradient boosting, support vector machines, or neural networks. GS: summary of fairness definitions, methodology for quantifying and visualizing fairness of credit risk scoring models, south German credit data, and simulation study.
KL: causal inference and couterfactual fairness. The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers.
Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher. The authors would like to thank the reviewers for their valuable feedback and acknowledge the support of the Institute of Applied Computer Science at Stralsund University of Applied Sciences for funding open access publication.
Agrawal, A. Google Scholar. Bischl, B. Editors M. Koster, P. Letmathe, R. Madlener, B. Peis, and G. Walther, 37—
0コメント