Model LineUpper: Supporting Interactive Model Comparison at Multiple Levels for AutoML

Title of the Paper

Model LineUpper: Supporting Interactive Model Comparison at Multiple Levels for AutoML

Paper Information

• Subject Area: Model comparison and interpretability in Automated Machine Learning (AutoML)
• Keywords: Automated Machine Learning, Automated Artificial Intelligence, Model Comparison, Explainable Artificial Intelligence (XAI), Visualization, Data Science, Performance Evaluation, Algorithm Transparency, Model Selection, Human-AI Collaboration

Research Background and Problems

• Identified Problems or Challenges:

  • Automated Machine Learning (AutoML) has made breakthroughs in simplifying the model development process, but existing AutoML systems provide limited support for model comparison, relying solely on performance metrics for selection.
  • In practice, users evaluate models based on more than just performance metrics, often considering model behavior on specific instances, error types, and feature importance.
  • Many models generated by AutoML are opaque "black boxes," leaving users without effective tools to understand their workings or compare them efficiently.

• Significance:

  • AutoML users require transparency and control to select the most suitable model based on domain knowledge and diverse criteria.
  • More effective model comparison tools can enhance data scientists' trust and engagement with AutoML results.

• Research Motivation and Related Work:

  • Inspired by Explainable Artificial Intelligence (XAI) technologies and model analysis tools, the authors aim to explore better ways to support model comparison in AutoML.
  • Existing research largely focuses on analyzing individual models rather than comparing multiple models, especially in scenarios where AutoML generates numerous candidates lacking adequate tool support.

Solution

• Proposed Method or Solution:

  • Developed a novel visualization tool, Model LineUpper, to support interactive comparison of candidate models generated by AutoML.
  • Integrated various XAI techniques to provide comparative information across multiple dimensions, from overall performance to specific instances.

• Innovations:

  • Offers three core views: a performance metrics table, a feature importance comparison view, and a probability scatter matrix, each supporting different levels of model comparison.
  • Enables users to examine overall feature importance, local feature importance for specific instances, and prediction probabilities for individual cases.
  • Uses visualization to help users identify key differences between models and understand the behavior of various algorithms.

• Implementation Steps and Key Technologies:

  • Performance Metrics Table: Displays commonly used metrics for binary classification tasks (e.g., F1 score, accuracy, ROC AUC) and highlights top-ranked metrics for easier comparison.
  • Feature Importance Comparison View: Visualizes global and local explanations using SHAP and SciKit-Learn feature importance values, allowing comparative analysis across different models.
  • Probability Scatter Matrix: Compares models' prediction probabilities for test data points, categorizing predictions into quadrants (correct/incorrect types) and supporting user selection of areas for deeper analysis.

Research Results

• Specific Outcomes:

  • Performance: User studies show that Model LineUpper significantly improves data scientists' efficiency in comparing AutoML models. Participants were able to select models based on different scenarios, such as reducing error rates or enhancing interpretability.
  • Feedback: Users reported increased understanding of model transparency and greater confidence in model selection. The System Usability Scale (SUS) score reached 3.98/5.
  • Support for Complex Thought Processes: The tool enables users to compare models not only based on performance metrics but also through multi-level detailed comparisons, including model errors, confidence distributions, and feature weights.

• Advantages:

  • Compared to existing AutoML systems, Model LineUpper focuses not only on model performance but also provides richer transparency and multi-dimensional comparison support.
  • Data scientists can select models based on specific business needs and preferences rather than relying solely on the "best model" recommended by AutoML systems.

• Experimental or Evaluation Results:

  • User ratings for view functionalities: global feature importance (average score 4.29/5), probability scatter matrix (4.21/5), local feature importance (3.93/5).
  • While completing tasks, users also suggested further improvements, such as more intuitive cross-feature sorting and optimized grouping for variant displays.

• Limitations and Future Directions:

  • Limitations:
    • Users showed lower utilization rates for certain complex views (e.g., local feature importance), indicating potential design optimization needs.
    • Visual challenges exist when comparing models generated by different algorithms (e.g., discrete vs. continuous probability distributions).
  • Future Directions:
    • Incorporate stronger tools for understanding data distributions to help users better interpret data.
    • Provide a more structured model comparison workflow to support multi-level exploration of algorithms and optimization methods.
    • Develop model comparison modules for more types of ML tasks (e.g., regression) to enhance tool versatility.

Conclusion

Model LineUpper is an innovative tool designed to address the challenges of model comparison in Automated Machine Learning. By integrating XAI technologies and intuitive visualization interfaces, it offers valuable insights to the research field and supports data scientists in model selection and decision-making. Future work can build upon its findings to further optimize tool design and better meet the practical needs of AutoML users.