Fraud Detection Ensemble Modeling for Extreme Class Imbalance

The patent combines supervised and unsupervised machine learning models into a hybrid ensemble system. The supervised models (e.g., XGBoost, Random Forest) leverage labeled fraud data while unsupervised models (e.g., Isolation Forest, Local Outlier Factor) detect anomalies without requiring labels. This merging allows the system to overcome the limitations of supervised learning alone in imbalanced datasets by incorporating unsupervised anomaly detection capabilities that do not suffer from class imbalance.

The patent creates a composite modeling approach by integrating multiple diverse models with different strengths. The ensemble includes supervised models trained on imbalanced data, unsupervised models for anomaly detection, and a meta-learning model that learns from the predictions of base models. This composite structure combines the advantages of different learning paradigms to achieve robust fraud detection despite extreme class imbalance.

The patent merges unsupervised models that avoid class imbalance issues with supervised models that provide accurate labeled predictions. The unsupervised models (Isolation Forest, Local Outlier Factor, Histogram-based Outlier Score) serve as base learners in the ensemble, contributing their ability to detect anomalies without requiring balanced labeled data. Their predictions are then integrated with supervised model predictions through the meta-learning model, combining the benefits of both approaches.

The patent introduces a meta-learning model as an intermediary that processes and integrates predictions from multiple base models (both supervised and unsupervised). The meta-model learns optimal ways to combine predictions from diverse sources, resolving the trade-off between unsupervised robustness to imbalance and supervised prediction accuracy by finding synergistic combinations of model outputs.

The patent segments the fraud detection task into multiple specialized models rather than using a single model. Different base models (supervised and unsupervised) are trained to detect different aspects of fraudulent behavior. The meta-learning model then segments the integration task by learning optimal weights and combinations for each base model's predictions, reducing false positives through diversified detection approaches.

The patent constructs a composite ensemble model that combines multiple diverse algorithms with different false positive characteristics. By integrating models like XGBoost, Random Forest, Isolation Forest, and Local Outlier Factor in a hybrid ensemble with meta-learning, the system creates a composite detection mechanism where the weaknesses of individual models are compensated by the strengths of others, significantly reducing overall false positive rates.

The patent uses a meta-learning model as an intermediary layer that manages the complexity of integrating multiple base models. The meta-model learns optimal strategies for combining predictions from diverse base models, automatically handling the complexity of ensemble integration. This intermediary structure allows the system to benefit from multiple models while the meta-learning component abstracts and manages the combinatorial complexity.

The patent dynamically adjusts parameters and weights of base models through meta-learning. The meta-model learns optimal weightings for each base model's predictions based on the specific characteristics of the data and problem instance. This parameter optimization allows the ensemble to achieve high accuracy while adapting the model configuration to minimize complexity for different fraud detection scenarios.