Model Implementation
This study applies both unsupervised and supervised learning algorithms to a real-world
financial transactions dataset to identify potentially fraudulent activity, particularly in contexts
such as money laundering and black-market finance. The dual-strategy approach addresses the
challenge posed by incomplete or delayed labeling in financial data, enabling both outlier
detection and explainable fraud classification.
Anomaly Detection using Isolation Forest
To complement traditional classification techniques and uncover novel fraud patterns, the
Isolation Forest algorithm was employed as an unsupervised anomaly detection method.
The model was trained using only two features—Amount (USD) and Money Laundering Risk
Score—selected for their known relevance in fraudulent behavior. A contamination rate of 1%
was chosen to flag the most significant outliers. The algorithm identified approximately 400
anomalous transactions, with a markedly high fraction of these also labeled as fraud by
supervised models, validating the model’s ability to detect high-risk outliers in an unlabeled
setting.
This methodology proved valuable in surfacing rare or evolving laundering tactics and
demonstrated potential for early warning systems, even in the absence of ground truth labels.
One-Class Support Vector Machine (SVM)
The One-Class SVM was applied as another unsupervised anomaly detection method. Using a
radial basis function (RBF) kernel with ν=0.05, the model was trained exclusively on the
feature space without class labels. The SVM learned the boundary enclosing the majority of the
data and labeled points outside this hypersphere as anomalous. After transforming the results to
binary fraud predictions, evaluation against pseudo-labels yielded competitive performance.
Although precision was limited (likely due to the tight boundary of SVM).
Autoencoder-Based Anomaly Detection
A deep autoencoder neural network was constructed to learn a compressed representation of
normal transaction behavior. The network architecture involved:
● An encoder reducing the input dimensionality,
● A symmetric decoder attempting to reconstruct the input.
Trained using Mean Squared Error (MSE) loss, the reconstruction error was used to flag
anomalies—transactions with high reconstruction errors (above the 95th percentile threshold)
were labeled as suspicious. This approach captured non-linear patterns effectively and
demonstrated high sensitivity to fraudulent behavior.
DBSCAN (Density-Based Spatial Clustering of Applications with Noise)
DBSCAN, a density-based clustering algorithm, was utilized to identify outlier points considered
noise in sparse regions of the feature space. Parameters eps=2.5 and min_samples=5 were
set based on empirical tuning. The algorithm labeled core, border, and noise points, with the
latter mapped to fraud predictions. While DBSCAN showed relatively lower accuracy, its
unsupervised nature provided a distinct clustering-based perspective and highlighted
transactions that diverged from any meaningful cluster.
Principal Component Analysis (PCA) for Outlier Detection
To capture linear structures and detect anomalies, PCA was employed for dimensionality
reduction followed by reconstruction. The first five principal components were retained, and
reconstruction error (MSE) between original and projected data was computed. Transactions
with reconstruction error above the 95th percentile were flagged. Although PCA is limited to
linear transformations, it performed surprisingly well, offering a lightweight baseline for anomaly
detection.
Comparative Performance Summary