Explanation Oriented Retrieval

1
Explanation Oriented Retrieval

• Dónal Doyle1, Pádraig Cunningham1, Derek Bridge2,
  Yusof Rahman1
• 1Department of Computer Science
• Trinity College Dublin, Ireland
• 2Computer Science, University College Cork, Cork,
  Ireland

2
A Fortiori arguments

• Why cant I go see Harry Potter and the Prisoner
  of Azkaban? My friend John saw it and hes 2
  years younger than me.

3
Outline

• Perceived Decision Boundary
• Similarity and Explanation Utility Measures
• Explanation Utility in Classification
• Evaluation
• Future Work

4
A Fortiori arguments

• Why cant I go see Harry Potter and the Prisoner
  of Azkaban? My friend John saw it and hes 2
  years younger than me.

5
Decision Boundary
Allowed See Harry Potter
Not Allowed See Harry Potter
Kate
John
Mark


Age
6
Explanation in CBR

• In CBR the most similar case is often used as the
  explanation for a classification
• However we believe that the nearest neighbour is
  not necessarily the most convincing case to
  explain a classification

7
Bronchiolitis child at AE

• 13 week old baby presented to an AE. Admit or
  Discharge?
• Compelling case 11 week baby that was
  discharged
• Less compelling case a 14 week old baby that was
  discharged

8
Decision Boundary
Admit
Discharge
14
11
13

Age

9
Similarity

• We use graphs to derive similarity scores from
  difference values

10
Similarity Graph for the feature Age
11
Utility Measure
Explainabilty ordering
Nearest Neighbours
Classification
Discharge
More about case 3 later
12
Explanation Utility Measure for feature Age
13
Explanation Utility Measure for feature Age
14
Blood Alcohol Domain

• Decision Surface based on
• Units of alcohol consumed
• Weight of individual
• Gender of individual
• Duration of drinking
• Food consumed

15
Sample Case
16
Sample Case
17
Sample Case
18
Sample Case
19
Sample Case
20
Utility Ranking for an Over the Limit Case
Weight (Kgs)
85
NUN
80
Case 1
Case 2
75
Q
70
7
8
9
10
11
12
13
Units Consumed
21
Explanation Utility Measures forClassification

• The utility metric for each outcome class is used
  to rank the entire case-base
• The utility score for the k nearest neighbours
  for each class is summed and the class with the
  highest score is returned as the prediction.

22
Classification Accuracy
23
Evaluation

• A problem in the diabetes domain was also
  evaluated
• An expert in the domain was presented with 9
  example cases
• In 8 out of the 9 cases the expert considered the
  case selected by the explanation utility more
  convincing than the nearest neighbour

24
Conclusions

• The nearest neighbour may not be the best case to
  explain a prediction
• The explanation utility neighbours attempts to
  select more convincing cases based on perceived
  decision boundary
• Expert evaluation agreed with our hypothesis in a
  medical domain

25
Future Work

• Evaluate the system in a multi-feature medical
  domain
• Develop textual explanations to accompany the
  case selected by the explanation utility measures
• Investigate the role of the NUN, which can be
  used as the rebuttal argument in the Toulmins
  Argument Structure

26
Gracias

• Questions?

27
Explanation Utility Measure for feature Age
28
Utility Ranking for an Over the Limit Case
Weight (Kgs)
85
NUN
80
Case 1
Case 2
75
Q
70
7
8
9
10
11
12
13
Units Consumed
29
(No Transcript)
30
Explanation Utility Measure