Enhancing Model Performance for Fraud Detection by Feature Engineering and Compact Unified Expressions

摘要

The performance of machine learning models can be improved in a variety of ways including segmentation, treating missing and outlier values, feature engineering, feature selection, multiple algorithms, algorithm tuning/ compactness and ensemble methods. Feature engineering and compactness of the model can have a significant impact on the algorithm's performance but usually requires detailed domain knowledge. Accuracy and compactness of machine learning models are equally important for optimal memory and storage needs. The research in this paper focuses on feature engineering and compactness of rulesets. Compactness of the ruleset can make the algorithm more efficient and derivation of new features makes the dataset high dimensional potentially resulting in higher accuracy. We have developed a technique to enhance model's performance with feature engineering and compact unified expressions for dataset of unknown domain using profile models approach. Classification accuracy is compared using well-known classifiers (Decision Tree, Ripple Down Rule and RandomForest). This technique is applied on fraud analysis bank dataset and multiple synthetic bank datasets. Empirical evaluation has shown that not only the ruleset size of training and prediction dataset is reduced but performance is also improved in other performance metrics including classification accuracy. In this paper, the transformed data is used for the experimental validation and development of fraud detection technique, but it can be used in other domains as well especially for scalable and distributed systems.