3. These
minions have
diabetes
Please check
the others, Dr.
Nefario
Learning with Positive and Unlabeled Data
Or….
We can use the data as is,
keeping in mind that the undiagnosed
minions might still have diabetes.
3
9. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
6
10. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
7
11. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
7
Tom
Age: 25
Sex: male
Known issues:
• Low vision
• Hot tibia
Jessa
Age: 27
Sex: female
Known issues:
• Lumbago
• Mono
Vincent
Age: 26
Sex: male
Known issues:
/
12. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
8
13. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
8
14. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
9
15. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
9
16. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
9
All have undesirable side effects
17. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
9
All have undesirable side effects
Complete database
18. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
9
All have undesirable side effects
Complete database
19. Positive and Unlabeled Data is Everywhere
Medical records Incomplete gene/protein databasesBookmarks/likes
10
20. What do I know?
• PhD Student @ Machine Learning Research Group, KU Leuven
• Fundamental research on learning with Positive and Unlabeled Data
• Estimating the Class Prior in Positive and Unlabeled Data through Decision
Tree Induction. AAAI, 2018.
• Positive and Unlabeled Relational Classification through Label Frequency
Estimation. ILP, 2017. (Most promising paper award)
• Ongoing work…
jessa.bekker@cs.kuleuven.be
people.cs.kuleuven.be/~jessa.bekker
53. Learn Naive Classifier, then Scale
Naive classifier predicts probability of being labeled
Scale
Option 1: So that proportion of positives is correct
=> Need to know proportion of positives!
Option 2: So that the maximum probability is 1
35
54. The extremely hard case:
Not separable but labels are
selected conditionally at random
36
55. Selected Conditionally At Random Assumption
Observed positive examples are
selected conditionally at random
from the positive set,
conditioned on the attributes.
The probability of a positive example to be selected
is a function of (some of) the attributes in the data,
called propensity score.
37
59. Learn Naive Classifier, then scale
Naive classifier predicts probability of being labeled
Scale
Use propensity score function
39
60. Learn Naive Classifier, then scale
Naive classifier predicts probability of being labeled
Scale
Use propensity score function
=> Need to know propensity score function!
39
61. Learn Classifier and Propensity Score
Simultaneously
Use available knowledge
• Attributes in propensity score function
• Proportion of positives
• Domain knowledge that classifier must adhere to
40
62. Learn Classifier and Propensity Score
Simultaneously
Use available knowledge
• Attributes in propensity score function
• Proportion of positives
• Domain knowledge that classifier must adhere to
40
63. Conclusions
• PU learning is very useful in practice
• We need assumptions to learn from PU data
• Linearly separable
• Selected completely at random
Scale probabilities
• Use proportion of positives
• Maximum scale
• Selected conditionally at random
Use propensity score
• Ongoing work
41