Exploring the Mind of Machine Learning with the What-If Tool

Summary:
Introducing the What-If Tool, a free application that helps you understand Machine Learning (ML) better. This blog post talks about what it can do, where you can use it in real life, and how it makes ML more transparent.

Unveiling a Breakthrough in ML Evaluation

Machine Learning (ML) systems have caused a big change, but it’s hard to know how well they work. The What-If Tool is a free tool that makes it easier. It helps you see how ML systems do in different situations, understand which data is important, and check if they treat everyone fairly. And you don’t need to be a coding expert to use it.

Exploring the Landscape of ML Understanding Frameworks and Flexible Visualization Platforms

The What-If Tool is different from other specialized tools. It’s like a versatile “black box” application. While tools like Uber’s Manifold and ModelTracker are also useful, the What-If Tool sets itself apart by emphasizing hypothetical testing, intersectional analysis, and fairness. It also introduces Facets Dive, a data exploration feature similar to Tableau and Pivot but made specifically for ML situations. Facets Dive keeps your data private while allowing you to explore it quickly.

Background and Overall Design

Understanding the What-If Tool’s design is crucial for comprehending its utility. This tool was crafted with the aim of making ML systems accessible to a broader audience, including data journalists, activists, and civil society groups. Fifteen months of rigorous testing and development, both internally and externally, culminated in a tool catering to users with varying degrees of ML experience.

Five core user needs shaped the tool’s functionality: testing multiple hypotheses without coding (N1), enhancing model understanding through visualizations (N2), conducting hypothetical testing without model access (N3), exploring intersectional analysis (N4), and evaluating performance improvements for multiple models (N5).

Functionalities of the What-If Tool (WIT)

WIT offers an array of features to analyze and interpret ML models:

· Data Exploration and Customizable Analysis: Users can interactively explore data and customize the analysis of data and model results.

· Feature Analysis: WIT provides summary statistics and distribution charts for all dataset features.

· Investigating What-If Hypotheses: Users can perform what-if analyses by editing data points and observing the effects on model inferences.

· Partial Dependence Plots: WIT offers partial dependence plots to understand how model predictions change with specific feature adjustments.

· Evaluating Performance and Fairness: Tools for analyzing model performance and fairness, including calculating performance metrics and applying fairness optimization strategies.

· Comparing Two Models: WIT supports the comparison of two machine learning models.

· Data Scaling: The tool’s capacity for handling data depends on factors such as the number of data points and the size of features.

Real-World Application

Three case studies that demonstrate WIT’s prowess in analyzing and interpreting ML models:

In the first case study, an ML researcher at a technology company utilized WIT to analyze a regression model used in production. The researcher wanted to understand how different features affected the model’s predictions, loaded a sample of data points, and manually edited feature values to observe how regression values changed. This led to the discovery of a bug in their ML pipeline, revealing an instance of training-serving skew that had gone undetected.

In the second case study, a software engineer at the same technology company used WIT to compare two versions of a regression model predicting a health metric for medical patients. WIT helped uncover a bug in the feature-building code for the first model, which had been causing incorrect predictions. This demonstrates WIT’s ability to identify issues that might not be apparent from traditional model evaluation metrics.

The third case study involved undergraduate students from MIT who used WIT to analyze the fairness of a linear classifier model simulating stop-and-frisk practices by the Boston Police Department. They used WIT to perform counterfactual reasoning, uncovering the unexpected importance of the police officer’s ID in predictions. The students also analyzed fairness across different groups and found disparities in the chances of being frisked based on factors like race and ethnicity.

Lessons from an Iterative Design Process

WIT’s development was an iterative process that aimed to streamline functionality and enhance user experience. It led to valuable design decisions that improved the tool’s usability.

Conclusion and Future Directions

To sum it up, the What-If Tool is a significant development in understanding ML models, suitable for both experts and beginners. It helps reveal hidden insights, solve mysteries, and provides clear transparency. In the future, there may be ways to make it even easier to use and understand ML models. The real power of understanding these models is in their ability to bring clarity and insight to a field that is constantly changing.

Reference: Wexler, J., Pushkarna, M., Bolukbasi, T., Wattenberg, M., Viégas, F., & Wilson, J. (2019). The What-If Tool: Interactive Probing of Machine Learning Models.