Voiceprint Clustering for Fraud Detection Runtime Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for identifying fraudsters in call centers face high runtime complexity, making it difficult to handle large volumes of calls and effectively cluster potential fraudsters, which can lead to inefficiencies in fraud detection.
Innovation Solution
A computerized method that processes voiceprints from interactions between callers and agents, calculates kth order moment scores, sorts and clusters voiceprints based on these scores, and performs searches to identify connected component voiceprints for fraud detection, reducing runtime complexity and improving fraudster clustering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If current systems process all voiceprints to identify fraudsters, then fraud detection accuracy is improved, but runtime complexity increases making the system unable to handle large volumes of calls
Solution Approach 1:
The patent segments the voiceprint processing task by dividing all voiceprints into multiple clusters based on similarity metrics. Each cluster represents a group of potentially related fraudster voiceprints. This segmentation allows the system to process clusters independently and in parallel, reducing overall runtime complexity while maintaining detection accuracy through comprehensive cluster analysis.
Solution Approach 2:
The patent performs preliminary clustering and filtering actions before the main fraud detection analysis. By pre-grouping voiceprints into clusters and identifying connected components in advance, the system reduces the computational burden during real-time fraud detection, enabling it to handle large volumes of calls efficiently.
2Reliability
If current systems process all voiceprints to cluster potential fraudsters, then clustering completeness is improved, but the system becomes unable to run due to high computational requirements
Solution Approach 1:
The patent divides the large-scale voiceprint dataset into manageable clusters based on similarity metrics. Each cluster contains voiceprints that are potentially related to the same fraudster. This segmentation enables the system to process and analyze clusters independently, maintaining clustering completeness while ensuring the system remains operational with large volumes of calls.
Solution Approach 2:
The patent implements a two-stage approach where it first performs a coarse clustering to identify potential fraudster groups, then applies more refined analysis only to promising clusters. This partial action strategy ensures comprehensive fraudster identification while keeping computational requirements within operational limits.
3Productivity
If the system reduces processing complexity to handle high call volumes, then productivity is improved, but the ability to accurately identify and cluster fraudsters deteriorates
Solution Approach 1:
The patent segments voiceprints into clusters based on similarity metrics, allowing efficient processing of high call volumes. Each cluster is then analyzed independently to identify connected components representing potential fraudsters. This segmentation maintains identification accuracy by ensuring thorough analysis within each cluster while enabling parallel processing for high productivity.
Solution Approach 2:
The patent replaces exhaustive brute-force comparison mechanisms with similarity-based clustering algorithms. This substitution reduces computational complexity from O(n²) to approximately O(n log n) or O(n), enabling the system to handle hundreds of millions of calls while maintaining fraudster identification accuracy through mathematically sound clustering approaches.
Data Source
AI summary
Methods for improved fraudster detection in a call center. A subset of a plurality of voiceprints from a plurality of interactions between callers and agents at a call center can be used as the basis for fraudster detection. A plurality of connected components that represents one or more voiceprints can be determined based on the subset of the plurality of voiceprints.