Collusion Detection System Using Graph Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods are inadequate for detecting collusion in click fraud, which involves multiple entities and evades detection by distributing fraudulent clicks across numerous sites and IP addresses, making it computationally expensive and time-consuming, and also struggle with predicting network traffic quality accurately due to constantly changing fraud schemes.
Innovation Solution
A comprehensive collusion detection and traffic quality prediction system that models fraud detection as graph partitioning and vector space clustering problems, using a modular, scalable, and distributed approach to identify and block bad traffic in real-time, incorporating a modeling module, toolkit, post processor, and job scheduler to analyze and filter out suspicious entities and patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional click fraud detection methods are used, then detection simplicity is maintained, but detection accuracy deteriorates due to inability to detect collusion across multiple entities
Solution Approach 1:
The patent segments the fraud detection problem into analyzing individual entities (IP addresses, referrers, publishers) and their pairwise relationships. By breaking down the complex collusion detection into smaller entity-level analyses, the system achieves higher detection accuracy without becoming computationally intractable. Each entity is scored independently based on its suspiciousness, and these scores are aggregated to detect collusive groups.
Solution Approach 2:
The patent introduces an intermediary scoring mechanism that assigns suspiciousness scores to individual entities based on their click patterns and relationships. This intermediary layer transforms the complex multi-entity collusion problem into a series of simpler entity-level scoring problems, enabling accurate fraud detection through aggregated entity scores rather than direct group analysis.
2Reliability
If fraud detection analyzes all possible entity combinations, then detection completeness improves, but computational cost increases exponentially
Solution Approach 1:
The patent divides the exhaustive combination analysis into independent entity-level assessments. Instead of evaluating all possible groups of entities, the system evaluates each entity separately and aggregates the results. This segmentation reduces computational complexity from exponential to linear or polynomial scale while maintaining detection reliability through the aggregation of individual entity suspiciousness scores.
Solution Approach 2:
Each entity independently evaluates its own suspiciousness based on its click patterns and relationships with other entities. The system allows entities to self-assess their fraud risk through automated scoring mechanisms, eliminating the need for centralized exhaustive analysis of all entity combinations. This self-service approach significantly reduces computational resource requirements.
3Adaptability or versatility
If detection systems use fixed rules, then system simplicity is maintained, but adaptability to new fraud schemes deteriorates
Solution Approach 1:
The patent implements dynamic detection by continuously updating entity suspiciousness scores based on evolving click patterns and relationships. Rather than relying on fixed rules, the system adapts to new fraud schemes by recalculating scores as new data arrives, allowing it to automatically respond to changing fraud tactics without requiring manual rule updates.
Solution Approach 2:
The system changes its detection parameters dynamically by adjusting entity scores based on observed behavior patterns. When new fraud schemes emerge, the scoring parameters automatically adapt to reflect the new patterns, enabling the system to maintain high adaptability while using a relatively simple scoring framework rather than complex rule sets.
4Productivity
If real-time fraud detection is implemented, then traffic quality improvement is achieved, but processing time requirements increase
Solution Approach 1:
The patent performs preliminary scoring of entities based on historical click data and established patterns before actual fraud detection is needed. By pre-computing entity suspiciousness scores and storing them for quick retrieval, the system enables rapid real-time detection without performing exhaustive analysis at the moment of decision, thus improving traffic quality in real-time while minimizing processing time loss.
Data Source
AI summary
Embodiments disclosed herein seamlessly integrate several components into a comprehensive collusion detection and traffic quality prediction system, including a strong modeling module for processing historical click data and transforming potential collusions hidden therein into solvable graph partitioning (network) and/or vector space clustering (pattern) models, a scalable and robust toolkit comprising a plurality of graph partitioning and clustering heuristics for analyzing and generating high density subgraphs and high dimensional clusters or groups, and a post processing module for extracting entities from the subgraphs and clusters and placing them on global block lists. Entities thus listed can be blocked from client networks in real time. As such, high traffic quality can be predicted. A job scheduler may schedule individual jobs from the modeling module based on the number of available resources in a distributed computing environment to minimize completion time while balancing load.