Document recommendation
Embodiments of the present disclosure provide a method and apparatus for document recommendation by obtaining a plurality of first data for a source recommendation task from activities related to a source user using a plurality of first documents; obtaining a plurality of second data for a target recommendation task from activities related to a target user using a plurality of second documents; performing the target recommendation task based on the plurality of first data, the plurality of second data, and knowledge transferred from the source recommendation task to obtain a target recommendation model; and conducting document recommendation to the target user using the target recommendation model.
1. A method for document recommendation within a data storage system, using one or more processors and memory, comprising:
obtaining a plurality of first data for a source recommendation task from activities related to a source user using a plurality of first documents;
obtaining a plurality of second data for a target recommendation task from activities related to a target user using a plurality of second documents;
performing the target recommendation task based on the plurality of first data, the plurality of second data, and information transferred from the source recommendation task to obtain a target recommendation model, wherein performing the target recommendation task comprises:
in response to obtaining the plurality of first data and the plurality of second data, extracting, from a plurality of eigenvectors, at least one eigenvector from the plurality of first documents and the plurality of second documents;
dividing at least a portion of respective data in the plurality of first data, the plurality of second data, and the at least one eigenvector extracted from the plurality of first documents and the plurality of second documents, into a plurality of subspaces with one or more variables for each subspace; and
optimizing the target recommendation task constrained by the one or more variables from one or more subspaces of the plurality of subspaces having equal importance in the plurality of first data and in the plurality of second data; and
conducting document recommendation to the target user using the target recommendation model.
2. The method according to claim 1 , wherein:
the source recommendation task comprises training a source prediction function (f S ) based on a source prediction vector (W S ) to obtain a source recommendation model; and
the target recommendation task comprises training a target prediction function (f T )based on a target prediction vector (W T ) to obtain the target recommendation model.
3. The method according to claim 2 , wherein performing the target recommendation task to obtain the target recommendation model further comprises:
obtaining a first optimization element that is a sum of differences between a plurality of prediction values obtained by performing recommendation prediction on the plurality of first documents using the source prediction vector and corresponding actual recommendation values, respectively;
obtaining a second optimization element that is a sum of differences between a plurality of prediction values obtained by performing recommendation prediction on the plurality of second documents using the target prediction vector and corresponding actual recommendation values, respectively;
obtaining a regularized element that is an element obtained by regularizing the source prediction vector or the target prediction vector under a condition that a modulus value of a sub-vector of the source prediction vector for each of the subspaces is equal to the modulus value of a sub-vector of the target prediction vector for that subspace; and
optimizing a sum of the first optimization element, the second optimization element, and the regularized element to obtain a value of the target vector corresponding to the scenario wherein the sum is minimal as a parameter for the target recommendation model.
4. The method according to claim 3 , further comprising:
generating, from the plurality of eigenvectors, a plurality of first eigenvectors (X S i ) based on the plurality of first data;
generating, from the plurality of eigenvectors, a plurality of second eigenvectors (X T i ) based on the plurality of second data; and
optimizing the expression:
min
w
S
,
w
T
,
a
∑
i
=
1
n
S
max
(
0
,
1
-
y
S
i
(
w
S
x
S
i
)
)
+
∑
i
=
1
n
T
max
(
0
,
1
-
y
T
i
(
w
T
x
T
i
)
)
+
λ
2
∑
k
=
1
c
(
a
k
)
2
s
.
t
.
∑
j
=
1
n
k
(
W
j
,
S
C
=
k
)
2
=
∑
j
=
1
n
k
(
W
j
,
T
C
=
k
)
2
=
(
a
k
)
2
,
k
=
1
,
2
,
…
,
c
wherein W S and W T denote the prediction vector for the source task and the prediction vector for the target task; n S denotes the number of the plurality of first eigenvectors; X S i denotes the i th first eigenvector; y S i denotes an actual recommendation value for the i th document; n T denotes the number of the plurality of second eigenvectors; X T i denotes the i th second eigenvector; Y T i denotes the actual recommendation value for the i th second document; λ is a constant; c denotes the number of the subspaces; n k denotes the number of features in a portion corresponding to the k th subspace in the first eigenvectors, which is identical to the number of features in a portion corresponding to the k th subspace in the second eigenvectors.
5. The method according claim 4 , wherein the source prediction function and the target prediction function are support vector machine linear classifiers.
6. The method according to claim 4 , wherein when the prediction vector w s denotes the prediction for a hyperplane of the target function , ƒ S ,then max (0,1−y S i (w S x S i ) denotes the difference between the predicted value for point X S i , in the source domain of the source task and the actual recommendation value y S i .
7. The method according to claim 4 , wherein d denotes the importance of the i th subspace variable in at least one of the source domain and target domain.
8. The method according to claim 1 , wherein the at least one eigenvector comprises one or more dimensions, wherein each of the one or more dimensions include a dimension value indicative of one or more keywords located in at least one of the plurality of first documents and the plurality of second documents.
9. The method according to claim 1 further comprising:
performing an offline analysis of the plurality of first data, the plurality of second data, and the information transferred from the source recommendation task.
10. The method according to claim 9 , wherein the transferred information comprises subspace importance of the plurality of first documents.
11. An apparatus for document recommendation on a data storage system, using one or more processors and memory, comprising:
an obtaining module, implemented via the one or more processors, configured to obtain a plurality of first data for a source recommendation task from activities related to a source user using a plurality of first documents, and to obtain a plurality of second data for a target recommendation task from activities related to a target user using a plurality of second documents;
a processing module, implemented via the one or more processors, configured to perform the target recommendation task based on the plurality of first data, the plurality of second data, and information transferred from the source recommendation task to obtain a target recommendation model, wherein the processing module is further configured to:
in response to obtaining the plurality of first data and the plurality of second data, extract, from a plurality of eigenvectors, at least one eigenvector from the plurality of first documents and the plurality of second documents;
divide at least a portion of respective data in the plurality of first data, the plurality of second data, and the at least one eigenvector extracted from the plurality of first documents and the plurality of second documents, into a plurality of subspaces with one or more variables for each subspace; and
optimize the target recommendation task constrained by the one or more variables from one or more subspaces of the plurality of subspaces having equal importance in the plurality of first data and in the plurality of second data; and
an outputting module configured to conduct document recommendation to the target user using the target recommendation model, and output a document recommendation result.
12. The apparatus according to claim 11 , wherein:
the source recommendation task comprises training a source prediction function (f s )based on a source prediction vector (W S )to obtain a source recommendation model; and
the target recommendation task comprises training a target prediction function (f T )based on a target prediction vector (W T )to obtain the target recommendation model.
13. The apparatus according to claim 12 , wherein the processing module is further configured to:
obtain a first optimization element that is a sum of differences between a plurality of prediction values obtained by performing recommendation prediction on the plurality of first documents using the source prediction vector and corresponding actual recommendation values, respectively;
obtain a second optimization element that is a sum of differences between a plurality of prediction values obtained by performing recommendation prediction on the plurality of second documents using the target prediction vector and corresponding actual recommendation values, respectively;
obtain a regularized element that is an element obtained by regularizing the source prediction vector or the target prediction vector under a condition that a modulus value of a sub-vector of the source prediction vector for each of the subspace is equal to the modulus value of a sub-vector of the target prediction vector for that subspace; and
optimize a sum of the first optimization element, the second optimization element, and the regularized element to obtain a value of the target vector corresponding to the scenario wherein the sum is minimal as a parameter for the target recommendation model.
14. The apparatus according to claim 13 , wherein the processing module is further configured to:
generate, from the plurality of eigenvectors, a plurality of first eigenvectors (X S i ) based on the plurality of first data;
generate, from the plurality of eigenvectors, a plurality of second eigenvectors (X T i ) based on the plurality of second data; and
optimize the expression:
min
w
S
,
w
T
,
a
∑
i
=
1
n
S
max
(
0
,
1
-
y
S
i
(
w
S
x
S
i
)
)
+
∑
i
=
1
n
T
max
(
0
,
1
-
y
T
i
(
w
T
x
T
i
)
)
+
λ
2
∑
k
=
1
c
(
a
k
)
2
s
.
t
.
∑
j
=
1
n
k
(
W
j
,
S
C
=
k
)
2
=
∑
j
=
1
n
k
(
W
j
,
T
C
=
k
)
2
=
(
a
k
)
2
,
k
=
1
,
2
,
…
,
c
wherein W S and W T denote the prediction vector for the source task and the prediction vector for the target task; n S denotes the number of the plurality of first eigenvectors; X S i denotes the ith first eigenvector; y S i denotes an actual recommendation value for the ith document; n T denotes the number of the plurality of second eigenvectors; X T i denotes the ith second eigenvector; Y T i denotes the actual recommendation value of the ith second document; λ is a constant; c denotes the number of the subspaces; n k denotes the number of features in a portion corresponding to the k th subspace in the first eigenvectors, which is identical to the number of features in a portion corresponding to the k th subspace in the second eigenvectors.
15. The apparatus according to claim 14 , wherein the source prediction function and the target prediction function are support vector machine linear classifiers.
16. The apparatus according to claim 11 , wherein the obtaining module, the processing module and the outputting module comprise at least one of a hardware element and a software application and a combination thereof.