17933907. DETECTING AND RECTIFYING MODEL DRIFT USING GOVERNANCE simplified abstract (International Business Machines Corporation)

DETECTING AND RECTIFYING MODEL DRIFT USING GOVERNANCE

Organization Name

International Business Machines Corporation

Inventor(s)

Neerju Gupta of Chelmsford MA (US)

Namit Kabra of Hyderabad (IN)

Yannick Saillet of Stuttgart (DE)

DETECTING AND RECTIFYING MODEL DRIFT USING GOVERNANCE - A simplified explanation of the abstract

This abstract first appeared for US patent application 17933907 titled 'DETECTING AND RECTIFYING MODEL DRIFT USING GOVERNANCE

Simplified Explanation

The embodiment described in the abstract is a system for monitoring machine learning models to detect and rectify model drift using governance. Here is a simplified explanation of the patent application:

• The system registers multiple machine learning models to a governance dashboard.
• It monitors these models to identify factors used by each model and group similar models into clusters.
• It automatically detects incorrect decisions made by a target model and calculates correlation scores with other models in the same cluster.
• If a correlation score is above a threshold, the system provides a cluster reinforcement recommendation.

Potential Applications

This technology could be applied in various industries where machine learning models are used, such as finance, healthcare, and marketing, to ensure the accuracy and reliability of these models over time.

Problems Solved

This technology addresses the issue of model drift, where machine learning models become less accurate over time due to changes in data or external factors. By monitoring and detecting drift, the system helps maintain the performance of these models.

Benefits

- Improved accuracy and reliability of machine learning models - Automated detection and rectification of model drift - Enhanced governance and oversight of machine learning processes

Potential Commercial Applications

The system could be valuable for companies that rely on machine learning models for decision-making, such as financial institutions, healthcare providers, and e-commerce platforms. The ability to monitor and rectify model drift can lead to better outcomes and increased trust in these models.

Possible Prior Art

One possible prior art in this field is the use of machine learning monitoring tools that track model performance and detect anomalies. However, the specific approach of grouping similar models into clusters and providing reinforcement recommendations based on correlation scores may be a novel aspect of this technology.

How does the system determine the threshold for correlation scores that trigger a reinforcement recommendation?

The abstract does not provide details on how the system sets the threshold for correlation scores. This could be an important factor in determining the effectiveness of the recommendations provided by the system.

What types of incorrect decisions can the system detect, and how does it differentiate between different types of errors?

The abstract mentions detecting incorrect decisions made by a target model, but it does not specify the types of errors that the system can identify. Understanding the range of errors that the system can detect and how it distinguishes between them could provide insights into the capabilities of this technology.

Original Abstract Submitted

An embodiment for monitoring machine learning models to detect and rectify model drift using governance. The embodiment may receive a plurality of machine learning models and register the plurality of machine learning models to a governance dashboard. The embodiment may automatically monitor the received plurality of machine learning models to identify factors used by each of the received plurality of machine learning models and generate corresponding clusters of similar machine learning models. The embodiment may automatically detect an incorrect decision made by a target machine learning model and then automatically calculate a correlation score between the target machine learning model and machine learning models within an associated corresponding cluster of similar machine learning models. The embodiment may, in response to detecting a correlation score above a threshold, automatically determine and output a cluster reinforcement recommendation.