Feature Selection Using Regularized Trees in Online Fraud Detection

• Activity Number: 656
• Type: Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Nitin Sharma*
• Companies: PayPal, Inc.
• Keywords: Feature Selection ; Random Forest ; Risk Analysis ; Fraud Detection ; Big Data ; Machine Learning

Abstract

Feature selection in online fraud detection poses several challenges due to high volume/high dimensionality in PayPal's highly diverse e-commerce and payment domain feature space. Filter methods involving orthogonal transformations or wrapper methods scale poorly requiring time/memory asymptotically exponential to dimensionality. Such methods require outlier treatment, data imputation, transformation and algorithm parameter tuning resulting into added feature selection complexities. Non-parametric methods of tree ensembles address these issues while rendering to distributed (e.g. map-reduce) frameworks. In addressing feature redundancy issues, we demonstrate a penalty-based regularization approach to a tree ensemble in an open-source big data machine-learning environment. Performance results using regularized, non-regularized tree features and selected filter/wrapper methods are presented. Applications of these regularized trees to generate highly predictive features or mini-models are introduced. Impact of tree regularization penalty functions on performance is quantified with specific recommendations on optimal parameterization for tree-based feature selection.