AI-POWERED INSIGHTS

Local Outlier Factor for Density-Based Anomalies

The Local Outlier Factor (LOF) algorithm identifies anomalies based on density deviations. Normal points reside in dense areas, while outliers are in sparse neighborhoods. LOF calculates a factor representing the density ratio of a point to its neighbors, with significantly higher values indicating potential outliers.

This method is particularly effective for "pseudo-normal" anomalies that are not globally extreme but appear isolated within their local context, addressing a common challenge in financial data analysis where local performance may seem out of place within a cluster of similar entities.

Single Class Support Vector Machine for Novel Anomaly Detection

Single Class Support Vector Machine (OCSVM) is adept at scenarios with only normal samples available for training. It constructs an optimal hypersphere or hyperplane to enclose the vast majority of normal data points, classifying any data outside this boundary as abnormal.

This is highly valuable for financial anomaly monitoring where labeled abnormal samples are scarce. OCSVM can detect new or unknown types of fraud by building a robust model of "normal behavior patterns" from abundant historical transaction data.

Advanced Ensemble and Deep Learning Models

The framework integrates Gradient Boosting Decision Trees (GBDT), specifically LightGBM, for efficient processing of high-dimensional structured financial data, learning from key indicators to identify abnormal records. Autoencoders perform unsupervised anomaly detection by identifying high reconstruction errors for abnormal data, enhanced by Transformer networks for temporal dependencies.

Graph Neural Networks (GNN) are employed to discover group association anomalies in financial networks. By modeling accounts/entities as nodes and transactions as edges, GNNs capture collaborative abnormal signals and hidden patterns of behavioral collusion.

Two-Stage Hybrid Monitoring Framework

The proposed framework operates in two stages:

1. Unsupervised Coarse Screening: Utilizes improved Isolation Forest (with time-dependent weights) and Stacked Autoencoders to quickly identify highly suspicious samples based on data distribution characteristics, without relying on prior labels.
2. Supervised Precision Judgment & Association Analysis: Leverages GBDT and Graph Neural Networks (GNNs). GBDT, augmented with techniques like SMOTE-ENN, refines anomaly classification using historical fraud cases and domain knowledge. GNNs analyze transaction networks to uncover group collusion and complex association anomalies.

Experimental Results and Performance Validation

Experiments were conducted on a mixed dataset combining public IEEE-CIS fraud detection data and synthetic data, specifically designed to simulate complex financial anomaly scenarios like multi-entity collusion. The framework was evaluated using key metrics suitable for imbalanced data, including Accuracy, Precision, Recall, F1-Score, AUC-ROC, and AUC-PR.

The complete framework achieved a Recall of 0.90, Precision of 0.87, and an F1-Score of 0.88, significantly outperforming benchmark models like Z-Score, Classic IForest, and AutoEncoder. Interpretability analysis using SHAP for LightGBM and visualization of GNN edges provided insights into decision drivers and critical trading paths.