Visualizations For Justifying Machine Learning Predictions David - - PowerPoint PPT Presentation

▶
visualizations for justifying machine learning predictions
SMART_READER_LITE
LIVE PREVIEW

Visualizations For Justifying Machine Learning Predictions David - - PowerPoint PPT Presentation

Mar 17, 2024 25 likes •146 views

Visualizations For Justifying Machine Learning Predictions David Johnson 1 Motivation Strengths of ML allowed expansion to diverse fields Fields and contexts far removed from traditional ML Users not trained in ML Eg. Medical

slide-1
SLIDE 1

Visualizations For Justifying Machine Learning Predictions

David Johnson

1

slide-2
SLIDE 2

Motivation

  • Strengths of ML allowed expansion to diverse fields
  • Fields and contexts far removed from traditional ML
  • Users not trained in ML
  • Eg. Medical field: Doctors use ML to predict disease given symptoms
  • The ML is a black box to them: Input → ? → Output

2

slide-3
SLIDE 3

Previous Work

Figure: Biran, O., MckKeown, K. (2014). Justification Narratives for Individual Classifications. AutoML workshop at ICML 2014.

3

slide-4
SLIDE 4

Previous Work

Some issues:

  • The vis relies on NLG quite a

bit

  • Vis isn’t very clear for

non-experts (what is Y-Prior? What is Slope?)

Figure: Biran, O., MckKeown, K. (2014). Justification Narratives for Individual Classifications. AutoML workshop at ICML 2014.

4

slide-5
SLIDE 5

Goals

  • Justify a ML prediction to a non-expert user
  • Show features providing evidence for/against the prediction
  • Select and visualize key features
  • Focus on interpretable models
  • Simplicity not complexity...

Figure: Munzner, T. (2014). Visualization Analysis and Design. CRC Press.

5

slide-6
SLIDE 6

Goals

  • Justify a ML prediction to a non-expert user
  • Show features providing evidence for/against the prediction
  • Select and visualize key features
  • Focus on interpretable models
  • Simplicity not complexity...

Figure: Munzner, T. (2014). Visualization Analysis and Design. CRC Press.

6

slide-7
SLIDE 7

Feature Visualizing

Vis can show effect and importance1

  • Effect: extent to which a feature contributes toward or against

prediction

  • Importance: Expected effect of the feature for a particular class

(mean feature value for the class)

1Biran, O., MckKeown, K. (2014). Justification Narratives for Individual Classifications. AutoML workshop at ICML 2014.

7

slide-8
SLIDE 8

Abstraction

  • Some raw data: arbitrary data with training/test sets
  • Task abstraction:
  • Analyze: discover, enjoy, derive
  • Data abstraction:
  • Items, attributes, values in a table
  • Two quantitative variables: effect, importance -- scatterplot

effective

8

slide-9
SLIDE 9

Demo

9

slide-10
SLIDE 10

Future Direction

NLG implemented Full web app implementation Expanded scope:

10

slide-11
SLIDE 11

Thanks!

Questions?

11

slide-12
SLIDE 12

12