Outlier Detection with Reinforcement Learning for Costly to Verify Data

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Outliers are common in datasets and require verification, which is time-consuming.
• Data errors and their causes can evolve, necessitating adaptive outlier detection.
• Existing methods may not optimally leverage verified outlier information.

Purpose of the Study:

• To develop an adaptive outlier detection approach using reinforcement learning.
• To enhance statistical outlier detection by dynamically tuning ensemble coefficients.
• To improve the efficiency and accuracy of identifying and managing data errors.

Main Methods:

• An ensemble of statistical outlier detection methods was employed.
• Reinforcement learning was applied to tune the ensemble's coefficients adaptively.
• The approach was validated using granular data from Dutch insurers and pension funds under Solvency II and FTK frameworks.

Main Results:

• The ensemble learner successfully identified outliers in the financial datasets.
• The integration of reinforcement learning further improved outlier detection performance.
• Adaptive coefficient optimization by the reinforcement learner enhanced overall accuracy.

Conclusions:

• Reinforcement learning offers an effective way to create adaptive outlier detection systems.
• The proposed method optimizes ensemble learning for improved identification of data errors.
• This approach provides a more efficient and accurate solution for dynamic outlier management.