Model interpretation - xAI

Allows users to perform a variety of analyses to interpret features and their contributions to the model's performance.

1. Setup analyses

Users can perform multiple analyses with select features to investigate their contributions to the model. Aside from display settings, which are explained on the side panel options page, users have two main settings for running these analyses:

• Vars: Selects the features to use in the analysis (if applicable).

• Analysis: Choose the analyses to perform on the selected features and generate plots for each.

Types of Analysis:

1. Scatter Plot: Visualize the relationship between two variables by displaying them as points distributed across a Cartesian plane. They help illustrate how changes in feature values could affect the model's output.

   • Select two variables to compare in this analysis

2. Heatmap: Depicts the interactions between two features by coloring cells according to the model’s output for combinations of feature values, showing how joint variations influence the prediction.

   • Select two variables for this analysis

3. ICE Plot (Individual Conditional Expectation): Visualize the change in the prediction outcome as a feature varies while all other features are held constant, highlighting the marginal effect of each feature

   1. Select one variable to vary in this analysis

4. LIME Plot (Local Interpretable Model-agnostic Explanations): Explain individual predictions by showing which features were most influential, helping to demystify the model’s behavior on a case-by-case basis.

   • Plots the most influential variables that best explain the individual predictions

   • Green bars indicate that a variable causes an increase in the probability (supports) the model. Red bars indicate that a variable causes a decrease in the probability (contradicts) the model.

   • Explanation Fit: Measure of how well the linear LIME model matches the region

5. Feature Importance: Illustrates the importance of each feature in the predictive model, showing which features have the most significant impact on the model's predictions.

6. Interactive Effects: Shows the interaction effects between features, highlighting how the combination of different features affects the model's output.

7. ALE Plot (Accumulated Local Effects): Visualize how features impact the prediction on average, focusing on how the model’s output changes when a feature varies over its distribution

8. PDP (Partial Dependence Plots) & ICE (Individual Conditional Expectation) Plot: Show both average and individual feature effects on the model's prediction

   1. Select variables to compare.

2. Display options

When a user runs analyses, they will be run for all the currently selected models and all the selected analyses. This is computationally expensive, so select only a few models and analyses at a time for best results. Once results are generated, the user can toggle which results to view.

• Select Model: At the top of the plot, select the model for which to view the analysis results.

• Select Analysis: On the right side-bar, select the analysis for which to view the resulting plot.