CNN Transaction Screening With Lossy Compression for SAR Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Enterprise organizations face challenges in accurately analyzing transaction history data across multiple channels due to the complexity of variables, leading to error-prone identification of suspicious activities, which results in false positives and inefficiencies in investigator review and resource allocation.
Innovation Solution
A computing platform uses lossy compression techniques and machine learning, specifically convolutional neural networks, to analyze transaction data in its entirety, combining it with customer profile data to generate accurate suspicious activity reports by determining probability scores and reducing false positives through vectorization and autoencoder methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If transaction data is approximated at an aggregate level for analysis, then the complexity of analyzing multiple channels is reduced, but the analysis becomes error-prone and loses accuracy
Solution Approach 1:
The patent segments the complex multi-channel transaction data into standardized feature vectors with consistent dimensions. Each transaction channel is broken down into discrete features (amount, frequency, counterparty, etc.) that are normalized and assembled into a uniform vector structure, enabling accurate analysis without aggregate approximation.
Solution Approach 2:
The patent transforms raw transaction data into standardized parameter representations through feature engineering. By changing the parameter format from diverse channel-specific structures to uniform feature vectors with consistent dimensions and scales, the system enables direct comparison and accurate analysis across all channels.
2Measurement precision
If machine learning models are trained on comprehensive transaction data, then identification accuracy improves, but computational resources and processing time increase
Solution Approach 1:
The patent performs preliminary feature extraction, normalization, and vectorization on transaction data before feeding it to the neural network. By pre-processing the data into standardized feature vectors and calculating key statistics in advance, the system reduces the computational burden during real-time analysis while maintaining high identification accuracy.
Solution Approach 2:
The patent extracts only the most relevant features from comprehensive transaction data, removing redundant information. By selecting and extracting key features (transaction amount, frequency, counterparty characteristics, etc.) and discarding unnecessary data, the system reduces processing time while preserving the accuracy needed for suspicious activity detection.
3Productivity
If false positive identifications are reduced through better analysis, then investigator efficiency increases, but more sophisticated analysis methods are required
Solution Approach 1:
The patent implements a feedback mechanism where the neural network's probability scores guide the investigation process. By providing confidence levels and probability distributions, the system allows investigators to focus on high-risk cases while automatically filtering low-risk transactions, improving efficiency without requiring overly complex manual analysis procedures.
Solution Approach 2:
The patent enables the analysis system to automatically perform feature extraction, normalization, and initial suspicious activity screening without investigator intervention. By making the system self-serve in preprocessing and initial analysis, investigators only need to review cases flagged by the algorithm, reducing their workload while maintaining high detection accuracy.
Data Source
AI summary
Aspects of the disclosure relate to using machine learning techniques for generating automated suspicious activity reports (SAR). A computing platform may generate a labelled transaction history dataset by combining historical transaction data with historical report information. The computing platform may train a convolutional neural network using the labelled transaction history dataset. The computing platform may receive new transaction data and compress the new transaction data using lossy compression. The computing platform may input the compressed transaction data into the convolutional neural network, which may cause the convolutional neural network to output a suspicious event probability score based on the compressed transaction data. The computing platform may determine whether the suspicious event probability score exceeds a predetermined threshold and, if so, the computing platform may send one or more commands directing a report processing system to generate a SAR, which may cause the report processing system to generate the SAR.