TL;DR
This paper introduces MonitorPlus, a novel surveillance system for parimutuel wagering that combines expert systems and machine learning to detect fraud, money laundering, and other illicit activities in real-time. By integrating a rule-based inference engine with machine learning for pattern discovery, MonitorPlus offers a flexible and adaptable solution for maintaining the integrity of a significant revenue source for governments.
What Problem Does This Paper Address?
Parimutuel wagering, a system where bettors collectively contribute to a pool from which winners are paid, is a substantial source of revenue for many state governments. However, this system is also vulnerable to various forms of fraud, crime, and money laundering. The challenge lies in effectively monitoring the vast amount of transaction data in real-time to detect suspicious activities without hindering legitimate wagering. Traditional methods often rely on manual reviews of historical data, which are slow, incomplete, and lack the ability to adapt to evolving criminal tactics. This paper addresses the need for a robust, automated, and intelligent surveillance system that can proactively identify and alert on potentially fraudulent activities, thereby protecting both the integrity of the wagering system and the revenue it generates.
Methods
MonitorPlus employs a hybrid artificial intelligence approach, combining a deductive inference engine with an inductive machine learning system.
- Expert System (Rule-Based): This component utilizes a business rule management system (Drools) to implement a knowledge base of predefined rules. These rules, often derived from subject matter experts (SMEs) and automatically generated from historical data, are designed to detect known patterns of fraudulent activity in real-time. The system uses a forward-chaining inference engine to fire these rules as new data becomes available.
- Machine Learning (Inductive): For discovering new and evolving suspicious patterns, MonitorPlus incorporates machine learning algorithms. Specifically, it uses decision trees (generated by C4.5) for automatic rule generation and a proprietary clustering algorithm. This allows the system to adapt its knowledge base over time and identify novel threats that might not be covered by existing rules.
- Data Processing: A "Tote Transaction Engine" validates and recomputes tote transactions, and a relational database management system supports online analytical processing (OLAP) for data analysis.
This multi-paradigm approach aims to provide both immediate detection of known threats and the ability to discover emerging ones.
Datasets
The paper does not explicitly detail specific datasets used for training or evaluation in terms of their origin or size. However, it mentions that the system was tested with:
- Actual parimutuel transaction data from two US states.
- Actual robotic wagering data from an ADW company.
- Millions of records of parimutuel wagering activity for stress testing.
- "What-if" queries and test files.
The C4.5 decision tree generation is noted to work best with training sets of around 100 to 500 records.
Key Insights and Ideas
- Hybrid AI is Crucial: The paper strongly advocates for a combined expert system and machine learning approach. While expert systems are good at detecting known patterns, machine learning is essential for discovering novel and evolving threats in a dynamic domain like parimutuel wagering.
- Real-time Alerts are Paramount: The core value of MonitorPlus lies in its ability to generate real-time alerts. This allows for immediate intervention, preventing potential fraud or money laundering from being completed, which is a significant advantage over post-facto analysis.
- Explainability Matters: The system's rule-based explanations are highlighted as critically important for legal action. This addresses a common challenge with black-box AI models, ensuring that alerts are not only generated but also understandable and defensible.
- Domain-Specific Challenges: The paper emphasizes the unique challenges of parimutuel wagering, such as the "over-round" in implied probabilities and the need to differentiate normal from abnormal betting behavior, which requires specialized surveillance techniques.
- Adaptability to Evolving Threats: Criminals are constantly innovating. MonitorPlus's ability to learn from new data and generate new rules is key to staying ahead of sophisticated fraud schemes.
Key Results
The paper focuses more on the system's architecture, capabilities, and the problems it addresses rather than presenting specific quantitative performance metrics (e.g., accuracy rates, false positive/negative rates). However, it highlights:
- Successful Validation: The first version of MonitorPlus was validated by RCI-IS and became the only tested and approved parimutuel surveillance system.
- Real-time Detection: The system successfully generates real-time alerts for various suspicious activities.
- Discovery of Suspicious Patterns: MonitorPlus has uncovered specific instances of:
- Pool arbitrage with a net rate of return of $1.50 per $1 wagered over 30+ days.
- Potential money laundering by grouping accounts with multiple small bets on favorites.
- Suspicious robotic wagering with unusually high returns.
- Automated Rule Generation: The C4.5 algorithm has been used to automatically generate rules for detecting potential money laundering.
Limitations and Future Work
- Limited Quantitative Results: The paper lacks detailed quantitative performance metrics for its detection capabilities.
- Dataset Specificity: While mentioning data sources, the paper doesn't provide specifics on dataset size, composition, or how they were split for training/testing, making it difficult to assess the robustness of the ML components.
- AI Technology Suitability: The authors note that while neural networks are powerful, they found them less suitable for this domain due to the need for massive training sets and difficulty in mapping weights to natural language explanations, which are crucial for regulators.
- Future Work: The paper mentions the ongoing evaluation of Drools 5 and the continuous development of the system. The authors believe the paradigm can be applied to other domains like lotteries, casinos, online gaming, and financial services. The development team continues to incorporate features and enhance the system.
Summary updated