Durham Reporter

Outlier detection is a critical task across various fields and industries. Traditional methods of univariate outlier detection can provide misleading results when analyzing outlier behavior. This is due to confounding factors that may make individuals or organizations appear normal while they are actually exhibiting anomalous behavior, and vice versa. Also, analyzing outliers based on a single outcome lacks the depth needed to fully capture human or organizational behavior complexity. This challenge is important in research and program administration, where outlier analysis is used to monitor behaviors, especially when financial awards or incentives are involved.

To address these issues, RTI International has introduced RTI Fovea, a generalized outlier behavior detection tool utilizing machine learning. It assists users in identifying multidimensional conditional outliers, enabling them to detect trends, limit malfeasance, learn from top performers, and make informed policy or investment decisions.

Fovea focuses on both accuracy and usability, providing a comprehensive approach to flagging anomalous behavior by accounting for confounding variables and allowing simultaneous analysis of multiple outcomes. For instance, an organization not identified as an outlier through traditional methods may still exhibit outlier characteristics when key confounders like geographical market and size are accounted for.

RTI Fovea employs a quantile forest version of the generalized random forest model proposed by Athey, Tibshirani, and Wager, with the flexibility to accommodate other predictive models. This model allows broad applications while maintaining computational efficiency. Fovea generates subject-specific predicted distributions for each signal variable and summarizes this with an Outlier Index to quantify outlier status relative to predicted distributions.

RTI Fovea includes built-in visualizations and additional outlier metrics, aiding users in identifying potential anomalous behavior. Users preferring more involvement can access model outputs for further analysis and visualization.

Despite its sophisticated methods, Fovea’s user-friendly design allows easy input of signal and control variables to produce outlier statistics and visualizations. This often requires less setup time compared to traditional outlier detection analyses.

Fovea's architecture and interface are designed to be highly generalizable, recognizing the wide range of potential use cases across industries. Examples include screening potential fraud by examining billing irregularities in healthcare, identifying unusual sensor data patterns for environmental monitoring, detecting outliers in student performance for targeted interventions, and analyzing employment data in workforce development.

RTI invites interested parties to learn more about how Fovea and other AI tools can assist in analyzing data efficiently and accurately.