Manifold: A Model-Agnostic Framework for Interpretation and Diagnosis of Machine Learning Models

Area of Science:

• Computer Science
• Machine Learning
• Data Visualization

Background:

• Interpretation and diagnosis of machine learning models are critical for trust and reliability.
• Existing techniques often lack generalizability, focusing on specific model architectures.
• A need exists for transparent and interactive tools applicable across diverse machine learning models.

Purpose of the Study:

• Introduce Manifold, a novel framework for visual analysis of machine learning models.
• Provide a generic approach for model interpretation, debugging, and comparison.
• Support complex scenarios involving integrated or heterogeneous model types.

Main Methods:

• Manifold utilizes visual analysis techniques, independent of internal model logic.
• The framework solely observes model inputs (instances, features) and outputs (predictions, probability distributions).
• Employs an iterative workflow: inspection (hypothesis), explanation (reasoning), and refinement (verification).

Main Results:

• Developed visual components including scatterplot-based summaries and customizable tabular views.
• Demonstrated Manifold's applicability to classification and regression tasks.
• Highlighted the framework's ability to facilitate transparent and interactive model diagnosis.

Conclusions:

• Manifold provides a versatile and generic solution for machine learning model interpretation and debugging.
• The visual approach enhances transparency and interactivity in the model development lifecycle.
• The framework shows potential for broad application across various machine learning use cases.