Active feature probing using data augmentation
View Patent ↗Systems and methods are disclosed that performs active feature probing using data augmentation. Active feature probing is a means of actively gathering information when the existing information is inadequate for decision making. The data augmentation technique generates factitious data which complete the existing information. Using the factitious data, the system is able to estimate the reliability of classification, and determine the most informative feature to probe, then gathers the additional information. The features are sequentially probed until the system has adequate information to make the decision.
1. A method to classify information, comprising:
augmenting existing information with factitious data;
performing actively feature probing using the factitious data;
generate a model based on the active feature probing with factitious data;
classifying information using the model; and
determining a goal for a given x obs as:
argmin
∈
un
E
Pr
(
x
l
❘
x
obs
)
(
Pr
(
❘
)
)
,
where ( ( )) is a loss function of a distribution ( ) on and Pr( | )= Pr( |x), obs+ =obs U { }, and un\ ={ ∈un| ≠ }.
2. The method of claim 1 , comprising learning existing information.
3. The method of claim 2 , comprising generating one or more feature vectors by clustering data from the existing information and the factitious data.
4. The method of claim 3 , comprising generating a classification model using the one or more feature vectors.
5. The method of claim 1 , learning information by clustering data from the existing information and the factitious data to generate one or more feature vectors and by generating a classification model from the one or more feature vectors.
6. The method of claim 1 , wherein the classifying information is done at run-time.
7. The method of claim 6 , comprising evaluating information in a new case to be classified.
8. The method of claim 7 , wherein the evaluating information comprises determining entropy of a set of factitious cases generated from the new case.
9. The method of claim 6 , comprising identifying the most similar past case as a solution.
10. The method of claim 1 , comprising gathering additional information for the new case.
11. The method of claim 1 , comprising determining a confidence on one or more outcomes after several probing.
12. The method of claim 11 , comprising measuring an entropy for the one or more outcomes with
H
(
y
|
x
obs
)
-
∑
y
∈
𝒴
Pr
(
y
|
x
obs
)
log
Pr
(
y
|
x
obs
)
.
13. The method of claim 1 , comprising determining a feature to probe.
14. The method of claim 1 , comprising finding a feature to minimize an expected loss.
15. The method of claim 1 , comprising obtaining the distribution of class label given all features, Pr(y|x), by a classification method.
16. The method of claim 15 , comprising sampling a number of augmented instances using a Monte Carlo method.
17. The method of claim 16 , comprising
approximating Pr(x i , |x obs ) without explicitly summing over all possible un\i.
18. The method of claim 1 , comprising drawing virtual instances.
19. The method of claim 18 , comprising obtaining samples from a distribution using a Gibbs sampling method.
20. The method of claim 18 , comprising obtaining samples from a distribution using a kernel density estimation.
21. A method to classify information, comprising:
augmenting existing information with factitious data;
performing actively feature probing using the factitious data;
generate a model based on the active feature probing with factitious data;
classifying information using the model; and
evaluating a probability by marginalizing a full probability
Pr
(
𝓍
𝒾
,
𝓎
|
x
obs
)
=
∑
x
un
\
𝒾
∈
𝒳
un
\
𝒾
Pr
(
𝓍
i
,
y
,
x
un
\
𝒾
|
x
obs
)
=
∑
x
un
\
𝒾
∈
𝒳
un
\
𝒾
Pr
(
𝓎
|
x
obs
,
x
un
\
𝒾
,
𝓍
i
)
Pr
(
x
un
\
𝒾
,
𝓍
i
|
x
obs
)
=
∑
x
un
\
𝒾
∈
𝒳
un
\
𝒾
Pr
(
𝓎
|
𝓍
)
Pr
(
x
un
|
x
obs
)
.
22. A method to classify information, comprising:
augmenting existing information with factitious data;
performing actively feature probing using the factitious data;
generate a model based on the active feature probing with factitious data;
classifying information using the model;
dynamically building a decision split based on sampled data from
Pr
(
x
un
|
x
obs
)
=
Pr
(
x
un
,
x
obs
)
/
Pr
(
x
obs
)
∝
1
Pr
(
x
obs
)
∑
X
∈
𝒟
K
(
x
,
z
)
∝
∑
𝒳
∈
𝒟
K
obs
(
x
obs
,
z
obs
)
∑
𝒾
∈
un
K
i
(
x
i
,
X
i
)
.