Learning classifiers for multiple-label data analysis
View Patent ↗A method for multiple-label data analysis includes: obtaining labeled data points from more than one labeler; building a classifier that maximizes a measure relating the data points, labels on the data points and a predicted output label; and assigning an output label to an input data point by using the classifier.
1. A method for multiple-label data analysis, comprising:
obtaining labeled data points from more than one labeler, wherein at least one data point is labeled by more than one labeler;
building a classifier that maximizes a measure relating the data points and labels on the data points to a predicted output label; and
assigning an output label to an input data point by using the classifier,
wherein the method is performed using a processor;
wherein building the classifier comprises solving the following equation
arg
min
θ
,
{
σ
t
}
∑
t
1
σ
t
2
∑
i
(
y
i
(
t
)
-
f
(
X
i
;
θ
)
)
2
+
log
σ
t
where y i (t) is a label for an i-th data point given by a t-th labeler and x i is the i-th data point and θ and σ are parameters of the classifier that are determined by solving the equation.
2. The method of claim 1 , wherein building the classifier comprises:
defining a functional form of a probability distribution that relates the labels on the data points to the predicted output label;
defining a functional form of a probability distribution that relates the predicted output label to the data points; and
maximizing a mutual information function by using the defined functional forms and the solution to the optimization problem to obtain a probability distribution over a plurality of possible predicted output labels given the input data point.
3. The method of claim 2 , wherein the probability distribution over the plurality of possible predicted output labels defines the classifier.
4. The method of claim 3 , wherein the classifier is deterministic.
5. The method of claim 1 , wherein the input data point to which the label is assigned is a newly obtained data point or a previously obtained data point.
6. The method of claim 1 , further comprising:
providing, in real-time, a user with the assigned output label; and
comparing the assigned output label to a label input to the data point by the user.
7. A computer program product comprising a non-transitory computer useable medium having computer program logic recorded thereon for multiple-label data analysis, the computer program logic comprising:
program code for obtaining labeled data points from more than one labeler, wherein at least one data point is labeled by more than one labeler;
program code for building a classifier that maximizes a measure relating the data points and labels on the data points to a predicted output label; and
program code for assigning an output label to an input data point by using the classifier,
wherein building the classifier comprises solving the following equation
arg
min
θ
,
[
σ
t
]
∑
t
1
σ
t
2
∑
i
(
y
i
(
t
)
-
f
(
x
i
;
θ
)
)
2
+
log
σ
t
where y i (t) is a label for an i-th data point given by a t-th labeler and x i is the i-th data point, and θ and σ are parameters of the classifier that are determined by solving the equation.
8. The computer program product of claim 7 , wherein the program code for building the classifier comprises:
program code for defining a functional form of a probability distribution that relates the labels on the data points to the predicted output label;
program code for defining a functional form of a probability distribution that relates the predicted output label to the data points; and
program code for maximizing a mutual information function by using the defined functional forms and the solution of the equation to obtain a probability distribution over a plurality of possible predicted output labels given the input data point.
9. The computer program product of claim 8 , wherein the probability distribution over the plurality of possible predicted output labels defines the classifier.
10. The computer program product of claim 9 , wherein the classifier is deterministic.
11. The computer program product of clam 7 , wherein the input data point to which the label is assigned is a newly obtained data point or a previously obtained data point.
12. The computer program product of claim 7 , further comprising:
program code for providing, in real-time, a user with the assigned output label; and
program code for enabling comparison, by the user, of the assigned output label to a label input to the data point by the user.
13. A method for analyzing data labeled by multiple-experts, comprising:
receiving a training dataset, wherein the training dataset includes labeled data points from expert-labelers, wherein at least one data point is labeled by more than one expert-labeler;
training a classifier that maximizes a measure relating a predicted label to the labels on the data points and the data points, wherein the classifier assigns a different weight to each expert based on expert-specific parameters; and
assigning an output label to an input data point by inputting the input data point to the classifier,
wherein the method is performed using a processor,
wherein building the classifier comprises solving by the following equation
arg
min
θ
,
[
σ
t
]
∑
t
1
σ
t
2
∑
i
(
y
i
(
t
)
-
f
(
x
i
;
θ
)
)
2
+
log
σ
t
where y i (t) is a label for an i-th data point given by a t-th labeler and x i is the i-th data point, and θ and σ are parameters of the classifier that are determined by solving the equation.
14. The method of claim 13 , further comprising:
grouping the experts based on the expert-specific parameters; and
assigning a different weight to a group of experts based on their grouping.
15. The method of claim 13 , further comprising:
retraining the classifier using newly provided labels or newly provided labeled data points.