ELIMINATE THE BLACK-BOX NATURE OF AI MODELS IN MULTIVARIATE DATA AND EXPLAIN THE FACTORS DECIDING THE RESULTS

Know More

CLIENT

One of the largest insurance providers in the world

BUSINESS CHALLENGES

The insurer had a manual triaging process to identify potentially fraudulent claims which comprised of rule-based engine to raise alerts and a large team to triage these alerts and select the candidate claims for further investigation. The objectives were as follows:
• Automate the triaging process with the AI model validating fraudulent claim alerts using multivariate data.
• Eliminate the black-box nature of the ML solution and explain the factors deciding the results for the claims.

SOLUTION

A machine learning solution was created to identify the patterns in the historical data and create triaging predictions on new incoming claims including:

Create features from the multivariate data along with the associated claims description (free text)
Create models with <0.2% false negatives using robust ML algorithms; none of the actual fraudulent claims was missed
Incorporate model explainability to improve the transparency of the decision making
Understand feature importance in predictions through global explanations, and provided what-if analysis by showing how predictions change with changes in inputs through local explanations

BUSINESS BENEFITS

Reduced the number of alerts to be reviewed by triage team by 90%, resulting in lowered cost of operations.

The explainability module indicated the important features influencing the predictions.

Ability to assess if features that should not impact predictions were doing so.

Improved the consistency of prediction leading to 98% accuracy against the human triaging, with no deviations in false negatives.