Fraud Detection Clustering for Transaction Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional fraud detection systems face challenges in efficiently classifying rare fraudulent events amidst a vast number of genuine transactions, often resulting in slow training procedures and models prone to false positives and false negatives.
Innovation Solution
A machine-learning system is developed that clusters transactions based on similarity measures, distinguishing between homogeneous and heterogeneous clusters. For heterogeneous clusters, specialized machine-learning models are trained, while homogeneous clusters use labeling information for classification, optimizing resource usage and improving classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single discriminative model is trained on the entire dataset, then the model can classify all transactions, but the training procedure becomes very slow and resource-intensive
Solution Approach 1:
The patent divides the entire training dataset into multiple clusters based on transaction similarity. Each cluster represents a subset of transactions with comparable characteristics. By segmenting the data this way, the system can train separate machine learning models on smaller, more manageable subsets rather than processing the entire large dataset with a single model, thereby reducing training time and computational resources while maintaining classification effectiveness.
2Reliability
If the model analyzes all training instances including easily classifiable genuine transactions, then comprehensive training is achieved, but the model becomes prone to false positives and false negatives
Solution Approach 1:
The patent extracts and identifies homogeneous clusters that contain predominantly one class (either mostly fraudulent or mostly genuine transactions). For these homogeneous clusters, the system applies simple majority voting based on cluster composition rather than running complex model inference. This extraction of easily classifiable transactions from the main processing flow reduces training time and prevents the model from overfitting to obvious patterns, thereby improving reliability by reducing false positives and false negatives.
3Adaptability or versatility
If rule-based detection logic is used, then the system is easy to implement and understand, but it is vulnerable to reverse engineering and cannot adapt to new fraud patterns
Solution Approach 1:
The patent implements a dynamic system where the fraud detection approach adapts based on the characteristics of each transaction cluster. The system automatically determines whether to apply simple majority voting or complex machine learning model inference depending on the homogeneity of the cluster. This dynamic adaptation allows the system to handle both simple and complex fraud patterns effectively, providing versatility without requiring a completely complex system architecture, as the complexity is applied selectively only where needed.
Data Source
AI summary
Embodiments of the present disclosure relate to a method, an apparatus and a computer program for providing a machine-learning system to be used for classifying transactions as either fraudulent or genuine, and a method, an apparatus and a computer program for classifying transactions as either fraudulent or genuine. The method for generating a machine-learning system for classifying transactions as either fraudulent or genuine based on a plurality of training transactions, each training transaction being associated with labelling information that indicates whether the training transaction is either genuine or fraudulent, comprises clustering the plurality of training transactions into a plurality of clusters based on a similarity measure. The method comprises determining, for each of the plurality of clusters, whether the cluster is homogeneous or heterogeneous. A heterogeneous cluster includes both fraudulent and genuine training transactions. A homogeneous cluster includes either only fraudulent or only genuine training transactions. The method comprises, for each of the heterogeneous clusters, training an associated machine-learning model using the training transactions belonging to the respective heterogeneous cluster.