Web Page Clustering for Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fraud detection techniques in website applications face challenges due to the large number of web pages and frequent changes, leading to high dimensionality in machine learning training data and reduced effectiveness over time, especially in detecting Account Take-Over (ATO) fraud.
Innovation Solution
A method for clustering web pages using a machine learning clustering process to group semantically related pages, constructing a directed graph of user browsing sessions, and applying a machine learning classifier to differentiate between legitimate and fraudulent activities, allowing for real-time detection and adaptation to changes in the website structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional fraud detection techniques are used with individual web page analysis, then detection coverage is maintained, but the system becomes computationally expensive and less effective due to high dimensionality
Solution Approach 1:
The patent merges multiple web page identifiers into cluster groups based on semantic similarity. Instead of analyzing each individual web page identifier separately, the system combines them into clusters that represent semantically related pages, thereby reducing the dimensionality of the data while preserving the essential information needed for fraud detection.
Solution Approach 2:
The patent segments the large space of web page identifiers into smaller, manageable clusters. By dividing the high-dimensional web page identifier space into multiple clusters, the system reduces computational complexity while maintaining detection accuracy through the use of cluster-level features in the fraud detection model.
2Adaptability or versatility
If the website structure changes frequently, then the application remains adaptable, but fraud detection effectiveness decreases due to outdated training data
Solution Approach 1:
The patent implements a dynamic clustering system that can adapt to changes in website structure. The clustering model is trained on historical data and can be retrained or updated as the website evolves, allowing the system to maintain effectiveness despite structural changes. The cluster representations capture semantic relationships that are more stable than individual page identifiers.
Solution Approach 2:
The patent transforms web page identifiers into cluster assignments as a new feature representation. This parameter transformation from individual page IDs to cluster labels creates a more stable and generalizable feature set that maintains detection effectiveness even when the underlying website structure changes, as clusters represent semantic categories rather than specific pages.
3Loss of information
If all web page identifiers are used for training, then comprehensive coverage is achieved, but training time and computational resources increase significantly
Solution Approach 1:
The patent combines multiple web page identifiers into cluster groups, reducing the number of features that need to be processed during training. By merging similar pages into clusters, the system preserves the informational content of individual pages while significantly reducing the computational burden of training the fraud detection model.
Solution Approach 2:
The patent extracts essential semantic information from individual web page identifiers and consolidates it into cluster representations. Instead of using all individual page identifiers for training, the system extracts the core semantic meaning and represents it through cluster assignments, thereby reducing training data volume while maintaining the essential browsing behavior information needed for fraud detection.
Data Source
AI summary
A machine learning clustering process is trained. Web pages of a website are clustered. User flow data associated with a first browsing session at the website is obtained. The user flow data includes a plurality of web page identifiers (e.g., URLs). A web page record for each of the web page identifiers is generated. Each web page record includes words of the corresponding web page identifier. Clusters of web page identifiers previously output from the trained machine learning clustering process are received. For each of the web page records, a cluster of web page identifiers is identified by mapping the web page record to one of the clusters of web page identifiers using the machine learning clustering process. A directed graph representative of the first browsing session is constructed. One or more nodes of the directed graph are the identified clusters of web page identifiers.