Case Study #1

 

Unstructured Data Mining Model

The Customer:  A Large Federal Government Agency

The Challenge:  Add Structure to Unstructured Data

  • Some government forms are not simple boxes, but require complex and unstructured responses.
  • This is especially true of legal filings.

The Solution: Utilize Text-Mining Problem Techniques

  • Using our “filing type” subject matter expertise, we identified examples of filings that could be used to create a gold dataset.
  • We created a custom XML format for representing the filing class with markup tags to identify key components.
  • We next hand-tagged the examples with the XML tags to create the gold dataset.
  • Using a variety of text-mining techniques, we found the keywords that represent each key component and applied the model to the full 1TB of filings.

The Results: Easily Accessible Data Available Quickly

  • Our and robust statistical model added XML tags to the entire group of filings.
  • Enabled investigators and regulators quicker access to vital information they needed to do their job.

 

Case Study #2

 

Workers Comp Fraud Detection

The Customer: United States Postal Service (USPS)

The Challenge: Detect Fraudulent and High Value Fraud Claims

  • Fraudulent claims are one of the largest avoidable expenses for the Office of Workers’ Compensation Programs
  • Develop a predictive model to:
    • Identify high-risk claims
    • Score claims based on their likelihood of being fraudulent
    • Identify key features that are predictive of fraudulent claims

The Solution: Regression Models and Predictive Performance

  • Data was analyzed for reliability, validity and completeness then classified into multiple sub-sets
  • Clustering and variable selection methods were used to predict outcomes.
  • Developed flexible and robust regression strategies to model (a) fraud and (b) amounts, providing thresholds that will raise flags for further investigation.
  • Assessed thresholds for predictive performance using split sample strategies.

The Results: Agency Expenses Reduced by Flagging Fraud Activity

  • Our flexible and robust statistical approach used clustering and ensemble methods to develop regression-based models for fraud detection and fraud value estimation
  • Models optimized for prediction error and thresholds using ROC methods to optimize false positive and false negative rates.

 

Case Study #3

 

Cross-Border Threat Model

The Customer: Large Federal Government Agency

The Challenge: Optimize Screenings and Reduce Threats

  • US Borders are nexus for smuggling drugs and human trafficking.
  • Tasked to develop predictive models for identifying and prioritizing screenings to reduce the threat of cross-border traffic.

The Solution: Predictive Models for Identifying Risks

  • Collected data to identify “bad actors” known to authorities
  • Collected data on those believed to be law-abiding transits
  • Developed metrics to quantify aspects of movement and social networks
  • Developed models to optimize predictions on whom and what was higher risk and should be prioritized for screenings

The Results: Increased Performance in Reduced Timeframes

  • Our flexible and robust statistical models were deployed to assist front line law enforcement officers in preparing inspections.
  • Law enforcement was able to perform fewer inspections and identify more contraband with the same staff in less time