Clustering-Based Subsampling for Fraud Detection Training Data

The patent applies local quality by treating different regions of the training data differently based on their fraud content. Clusters containing fraudulent transactions are identified and preserved with higher fidelity, while clusters without fraud undergo more aggressive subsampling. This localized differential treatment ensures that representative fraudulent information is retained while still achieving training speed improvement through subsampling of the majority genuine transactions.

The patent changes the subsampling parameter (sampling rate) based on the fraud content of each cluster. Instead of applying a uniform subsampling rate to all data, the method dynamically adjusts the sampling parameter: clusters with fraud transactions use a higher sampling rate to preserve representative instances, while clusters without fraud use a lower sampling rate to reduce data size. This parameter adaptation resolves the contradiction between speed and information retention.

The patent implements local quality by applying different subsampling strategies to different clusters based on their fraud characteristics. Clusters containing fraudulent transactions are identified through clustering analysis, and these specific regions are preserved with higher sampling rates to maintain fraud detection accuracy. Meanwhile, clusters consisting only of genuine transactions are more aggressively subsampled to reduce overall data size. This localized differential subsampling resolves the contradiction between reducing data quantity and maintaining detection precision.

The patent segments the training data into multiple clusters based on transaction characteristics and fraud content. By dividing the homogeneous training set into heterogeneous clusters, the method can apply targeted subsampling to each segment. This segmentation allows the system to reduce overall data size while preserving the critical fraud-containing segments, thereby maintaining detection accuracy despite reduced data quantity.

The patent extracts and separates the essential fraudulent information from the complete training set through clustering analysis. By identifying clusters that contain fraudulent transactions and extracting these critical segments, the method preserves the core information needed for accurate classification. The remaining non-essential genuine transactions are then subsampled or removed, significantly reducing training time while maintaining classification accuracy through the preserved fraud-containing clusters.

The patent performs preliminary clustering analysis before the actual training process to identify and separate fraud-containing clusters. This preliminary action allows the system to pre-determine which data segments are critical for accuracy and which can be reduced. By performing this classification and segmentation beforehand, the method enables efficient training on a reduced dataset while ensuring that the essential fraud patterns are already identified and preserved for accurate classification.