سجاد عليزاده فرد

Fraud in financial data is always a serious concern for business organizations and individuals. Applying manual checks to detect fraud is time-consuming and expensive. Proper fraud detection allows investigators to take timely action and prevent further fraud and financial losses. One of the main step in the fraud detection process is the feature selection step, which has an important impact on the accuracy and execution time of the models. Credit card transactions typically have many features. Some features may not be meaningful to the classifiers or lead to overfitting. In addition, having fewer duplicated features leads to a better understanding of the classifier decision. Also, feature selection can increase the "speed of classifiers" due to reducing the size of the feature set and "classifier performance" due to avoiding overfitting. One of the main challenges when using complex models in fraud detection is the lack of “explainability” about how the models work and why they make decisions. In particular, when working with sensitive data in security domains, providing effective explanations to system users is of great importance and has become an ethical and legal requirement in many applied fields. In this research, we present "an explainable feature selection framework" based on ensemble approach. In this work, we applyed the proposed framework on different combinations of best models in the previous works and compared its results with other feature selection algorithms quantitatively and qualitatively. Quantitative eva‎luation of the "X-SHAoLIM" framework on different combinations of best models in previous works showed that the use of the proposed framework in the feature selection step, on average, increases the accuracy of the models, based on precision(+5.6), recall (+ 1.5), F-Score (+3.5) and AUC (+6.75) and compared to other feature selection algorithms, it brings the best performance. In addition to increasing the accuracy of the models, the proposed framework, due to the use of explainable algorithms like SHAP and LIME, has more capability in interpretibility and analysis of features importance on the model predictions and provides effective explanations to the system users.

كشف تقلب , يادگيري ماشين , انتخاب ويژگي‌ها , يادگيري گروهي , تفسيرپذيري , داده‌كاوي

fraud detection , machine learning , feature selection , ensemble learning , explainability , data mining

sajjad alizadeh fard

hossein rahmani