Multi-label classification for overlapping classes
In an example, there is disclosed a computing apparatus, including one or more logic elements comprising a classification engine operable to identify at least one overlapping class pair comprising at least two classes selected from a class group, wherein the overlapping class pair may be characterized by a substantial mutual false positive rate; and assign an object to at least one assigned class selected from the class group.
1. A computing apparatus, comprising:
one or more logic elements comprising a classification engine operable to:
build a universal hyperclassifier;
identify at least one overlapping class pair comprising at least two classes selected from a class group, wherein the overlapping class pair may be characterized by an at least approximately 50% mutual false positive rate; and
assign an object to at least one assigned class selected from the class group.
2. The computing apparatus of claim 1 , wherein the at least one assigned class is selected from the overlapping class pair.
3. The computing apparatus of claim 1 , wherein the classification engine is further operable to build at least one ensemble learning model, comprising at least two binomial experts.
4. The computing apparatus of claim 3 , wherein the classification engine is further operable to:
assign the object a first match probability index for a first class;
assign the object a second match probability index for a second class;
determine that the difference between the first match probability index and second match probability index is less than a threshold α; and
invoke the at least one ensemble learning model comprising applying a first binomial expert and a second binomial expert to the first and second classes.
5. The computing apparatus of claim 3 , wherein the classification engine is further operable to compound predictions of the at least two binomial experts.
6. The computing apparatus of claim 1 , wherein building a universal hyper-classifier comprises mapping an instance i of the object to a plurality of classes that can be associated with that instance.
7. The computing apparatus of claim 1 , wherein the classification engine is further operable to build a multinominal classifier.
8. The computing apparatus of claim 7 , wherein building the multinomial classifier comprises an L2 regularized logistic regression model of the form:
min
w
λ
2
w
2
+
Σ
i
=
1
m
log
(
1
+
e
-
y
i
w
T
x
i
)
.
9. The computing apparatus of claim 1 , wherein identifying at least one overlapping class pair comprises determining that a false positive rate exceeds a threshold T.
10. The computing apparatus of claim 1 , wherein identifying at least one overlapping class pair comprises building a confusion matrix.
11. The computing apparatus of claim 1 , wherein the object is a text object.
12. The computing apparatus of claim 1 , wherein identifying overlapping class pairs comprises classifying a model according to:
min
w
λ
2
w
2
+
Σ
i
=
1
m
(
max
(
0
,
1
-
y
i
w
T
x
i
)
)
.
13. The computing apparatus of claim 1 , wherein identifying overlapping class pairs comprises classifying a model according to:
min
w
λ
2
w
2
+
Σ
i
=
1
m
(
max
(
0
,
1
-
y
i
w
T
x
i
)
)
.
14. One or more tangible, non-transitory computer-readable mediums having stored thereon executable instructions for providing a classification engine operable to:
build a universal hyperclassifier;
identify at least one overlapping class pair comprising at least two classes selected from a class group, wherein the overlapping class pair may be characterized by an at least approximately 50% mutual false positive rate; and
assign an object to at least one assigned class selected from the class group.
15. The one or more tangible, non-transitory computer readable mediums of claim 14 , wherein the at least one assigned class is selected from the overlapping class pair.
16. The one or more tangible, non-transitory computer readable mediums of claim 14 , wherein the classification engine is further operable to build at least one ensemble learning model, comprising at least two binomial experts.
17. The one or more tangible, non-transitory computer readable mediums of claim 16 , wherein the classification engine is further operable to:
assign the object a first match probability index for a first class;
assign the object a second match probability index for a second class;
determine that the difference between the first match probability index and second match probability index is less than a threshold α; and
invoke the at least one ensemble learning model comprising applying a first binomial expert and a second binomial expert to the first and second classes.
18. The one or more tangible, non-transitory computer readable mediums of claim 16 , wherein the classification engine is further operable to compound predictions of the at least two binomial experts.
19. The one or more tangible, non-transitory computer readable mediums of claim 14 , wherein building a universal hyper-classifier comprises mapping an instance i of the object to a plurality of classes that can be associated with that instance.
20. The one or more tangible, non-transitory computer readable mediums of claim 14 , wherein the classification engine is further operable to build a multinominal classifier.
21. The one or more tangible, non-transitory computer readable mediums of claim 20 , wherein building the multinomial classifier comprises an L2 regularized logistic regression model of the form:
min
w
λ
2
w
2
+
Σ
i
=
1
m
log
(
1
+
e
-
y
i
w
T
x
i
)
.
22. A computer-implemented method of providing a classification engine, comprising:
building a universal hyperclassifier;
identifying at least one overlapping class pair comprising at least two classes selected from a class group, wherein the overlapping class pair may be characterized by an at least approximately 50% mutual false positive rate; and
assigning an object to at least one assigned class selected from the class group.
23. The computer-implemented method of claim 22 , wherein the at least one assigned class is selected from the overlapping class pair.
24. The computer-implemented method of claim 22 , wherein the classification engine is further operable to build at least one ensemble learning model, comprising at least two binomial experts.
25. The computer-implemented method of claim 24 , wherein the classification engine is further operable to:
assign the object a first match probability index for a first class;
assign the object a second match probability index for a second class;
determine that the difference between the first match probability index and second match probability index is less than a threshold α; and
invoke the at least one ensemble learning model comprising applying a first binomial expert and a second binomial expert to the first and second classes.