StackGenVis: Alignment of Data, Algorithms, and Models for Stacking Ensemble Learning Using Performance Metrics.

Area of Science:

• Machine Learning
• Data Visualization
• Artificial Intelligence

Background:

• Ensemble methods like stacking enhance machine learning predictive performance by combining diverse models.
• Creating effective stacked generalization models from scratch is complex due to vast parameter and feature spaces.
• Existing methods lack efficient tools for navigating the solution space and optimizing stacked models.

Purpose of the Study:

• To present StackGenVis, a novel visual analytics system designed to support stacked generalization in machine learning.
• To assist users in dynamically managing data, selecting features, and choosing optimal base models for ensemble creation.
• To reduce the complexity of stacked models by identifying and removing underperforming or overpromising components.

Main Methods:

• Development of a knowledge generation model integrated with visualization techniques for ensemble learning.
• Implementation of the StackGenVis system enabling dynamic adaptation of performance metrics and feature selection.
• Utilizing user interaction for managing data instances and selecting diverse, high-performing algorithms.

Main Results:

• StackGenVis facilitates informed decision-making in model selection and stack simplification.
• The system effectively aids in identifying top-performing and diverse algorithms for ensemble construction.
• Demonstrated applicability on real-world healthcare and text sentiment/stance detection datasets.

Conclusions:

• StackGenVis significantly simplifies the process of building stacked generalization models.
• The visual analytics approach enhances user control and understanding in complex machine learning tasks.
• Expert evaluations confirm the tool's effectiveness in managing and optimizing ensemble models.