Path7.Mod1.c - Responsible AI Dashboard - Evaluate the RAI Dashboard Flashcards

1
Q

EA Exp Cfs CA

Depending on the Components selected, the Insights you’ll see on your dashboard

A
  • Error Analysis
  • Explanations
  • Counterfactuals
  • Causal Analysis
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Describe Error Analysis

A

You can review and understand how Errors are distributed in your dataset i.e. are there more errors with cohorts representing Women vs Men, Blacks vs Whites, when the person is disabled, etc.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

ETM EHM

Two visual representations for Error Analysis

A

Error Tree Map: Explore which combinations of cohorts results in more false predictions

Error Heat Map: Grid overview of a model’s errors over the scale of selected features (ie. the relationship of features and their values and the % error yield per combination)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

The dumb name she gave the entity from that Haunted episode…

  • Describe Explanations w.r.t. Model Explainers
  • Name the most common explainer used and what it does
A
  • When you want to understand and calculate Feature Importance ie how much a feature influenced the final prediction. You can use various Model Explainers to calculate it.
  • The most common explainer is the mimic explainer. It trains an interpretable model using the same data and task.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

AFI IFI

Two types of Feature Importance available in the Feature Importance Bar Chart

A

Aggregate Feature Importance: How each feature in the test data influences the model’s prediction overall

Individual Feature Importance: How each feature in the test data impacts an individual prediction. When selected, you are presented a grid of all individual data points that were scored. Select one or many and you’ll see to what degree the features influenced those datapoint’s result

In other words, Overall influence vs Individual prediction influence

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

Marvel’s….

Describe Counterfactuals

A

What-If scenarios i.e. how a model’s output changes based on a change in the input, that provide deeper insight into feature importance.

Select a datapoint and the desired model’s prediction for that point. Create a What-If Counterfactual in the panel to discover your desired prediction for differing input.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

“Mean Feature Importance”

Describe Causal Analysis

A

The use of statistical techniques to estimate the average effect of a feature on a desired prediction.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

ACE ICE TP

Three available Causal Analysis tabs in the RAI Dashboard

Describe what a Treatment Feature is

A
  • Aggregate Causal Effects: Average Causal Effects for predefined Treatment Features
  • Individual Causal Effects: Change individual Treatment Features to explore their influence on a prediction
  • Treatment Policy: What parts of your data points benefit most from Treatment

Treatment Features - Features you want to change to optimize predictions

How well did you know this?
1
Not at all
2
3
4
5
Perfectly