Relevance table aggregation in a database system for providing video recommendations
View Patent ↗In one embodiment, a method generates a plurality of sub-relevance tables including a first set of relevance values between media programs. Each table models relevance values for a single feature in a plurality of features. Labeling results are received that include a second set of relevance values between the media programs. The method combines the sub-relevance tables into a single relevance table that includes a third set of relevance values between the media programs for the plurality of features. The combining generates weights for each of the sub-relevance tables based on the second set of relevance values for the labeling results and the first set of relevance values of the sub-relevance tables that are used to generate the third set of relevance values. A recommendation is provided to a user using the third set of relevance values from the single relevance table and a characteristic of the user.
1. A method comprising:
generating, by a computing device, a plurality of sub-relevance tables including a first set of relevance values between a first set of media programs and a second set of media programs being provided by a video delivery service, wherein each table models relevance values relating similarity for a single feature in a plurality of features between the first set of media programs and the second set of media programs;
receiving, by the computing device, labeling results including a second set of relevance values between the first set of media programs and the second set of media programs, the labeling results based on user input;
generating, by the computing device, weights for each of the plurality of sub-relevance tables based on the second set of relevance values for the labeling results and the first set of relevance values of the plurality of sub-relevance tables;
combining, by the computing device, relevance values from the plurality of sub-relevance tables into a single relevance table by applying respective weights for each of the plurality of sub-relevance tables to the relevance values, wherein the single relevance table includes a third set of relevance values relating similarity for the plurality of features between the first set of media programs and the second set of media programs; and
providing, by the computing device, a recommendation to a user using the video delivery service, the recommendation generated using the third set of relevance values from the single relevance table and a characteristic of the user.
2. The method of claim 1 , further comprising:
analyzing at least one sub-relevance table in the plurality of sub-relevance tables to generate a plurality of questions regarding relationships between media programs in the first set of media programs and the second set of media programs; and
receiving the user input answering the plurality of questions, the user input used to generate the labeling results.
3. The method of claim 2 , wherein a question in the plurality of questions comprises tuples including a relationship between three media programs.
4. The method of claim 3 , wherein:
the question asked whether a first media program or a second media program is more relevant to a third media program, and
a selection of the first media program or the second media program is used to generate a labeling result.
5. The method of claim 1 , wherein the weights are generated by using a function that measures a correctness of a sub-relevance table by comparing the relevance values for the sub-relevance table to the labeling results.
6. The method of claim 5 , wherein the function comprises:
L
=
∑
(
k
,
i
,
j
,
)
∈
P
1
1
+
exp
(
α
(
F
(
x
k
,
i
)
-
F
(
x
,
j
)
)
,
where (k, i, j) and P are the labeling results; x k,i is a vector containing the relevance values from media program k to media program i in the sub-relevance tables, x k,j is a vector containing the relevance values from media program k to media program j in the sub-relevance tables s; α is a scaling parameter, and F( )is a scoring function with respect to the single relevance table.
7. The method of claim 1 , wherein combining the plurality of sub-relevance tables into the single relevance table comprises using a linear combination to combine the weights for each of the plurality of sub-relevance tables.
8. The method of claim 7 , wherein the linear combination uses:
F ( x )= w T x,
where w is a weight vector and x is the vector containing the relevance values between the first set of media programs and the second set of media programs.
9. The method of claim 1 , wherein combining relevance values from the plurality of sub-relevance tables into the single relevance table comprises using a non-linear combination to combine the weights for each of the plurality of sub-relevance tables.
10. The method of claim 9 , wherein the non-linear combination uses:
F
(
x
)
=
lim
t
→
∞
F
t
(
x
)
=
∑
k
=
1
t
β
k
f
k
(
x
)
,
where F t (x) is a scoring function in the t-th iteration with respect to the single relevance table and F(x) is a final scoring function with respect to the single relevance table and β k is the weight of F k(x) .
11. The method of claim 1 , further comprising:
adjusting the single relevance table based on feedback information from users.
12. The method of claim 11 , wherein adjusting comprises:
providing a set of recommendations to the user using the single relevance table; and
receiving the feedback information based on a response from the users for the set of recommendations.
13. The method of claim 1 , wherein combining relevance values from the plurality of sub-relevance tables into the single relevance table comprises modeling a tie in the labeling results using a function that models the tie.
14. A non-transitory computer-readable storage medium containing instructions, that when executed, control a computer system to be configured for:
generating a plurality of sub-relevance tables including a first set of relevance values between a first set of media programs and a second set of media programs being provided by a video delivery service, wherein each table models relevance values relating similarity for a single feature in a plurality of features between the first set of media programs and the second set of media programs;
receiving labeling results including a second set of relevance values between the first set of media programs and the second set of media programs, the labeling results based on user input;
generating weights for each of the plurality of sub-relevance tables based on the second set of relevance values for the labeling results and the first set of relevance values of the plurality of sub-relevance tables;
combining relevance values from the plurality of sub-relevance tables into a single relevance table by applying respective weights for each of the plurality of sub-relevance tables to the relevance values, wherein the single relevance table includes a third set of relevance values relating similarity for the plurality of features between the first set of media programs and the second set of media programs; and
providing a recommendation to a user using the video delivery service, the recommendation generated using the third set of relevance values from the single relevance table and a characteristic of the user.
15. The non-transitory computer-readable storage medium of claim 14 , further configured for:
analyzing at least one sub-relevance table in the plurality of sub-relevance tables to generate a plurality of questions regarding relationships between media programs in the first set of media programs and the second set of media programs; and
receiving the user input answering the plurality of questions, the user input used to generate the labeling results.
16. The non-transitory computer-readable storage medium of claim 14 , wherein the weights are generated by using a function that measures a correctness of a sub-relevance table by comparing the relevance values for the sub-relevance table to the labeling results.
17. The non-transitory computer-readable storage medium of claim 16 , wherein the function comprises:
L
=
∑
(
k
,
i
,
j
,
)
∈
P
1
1
+
exp
(
α
(
F
(
x
k
,
i
)
-
F
(
x
,
j
)
)
,
where (k, i, j) and P are the labeling results; x k,i is a vector containing the relevance values from media program k to media program i in the sub-relevance tables, x k,j is a vector containing the relevance values from media program k to media program j in the sub-relevance tables s; α is a scaling parameter, and F( )is a scoring function with respect to the single relevance table.
18. The non-transitory computer-readable storage medium of claim 14 , wherein combining comprises using a linear combination of:
F ( x )= w T x,
where w is a weight vector and x is the vector containing the relevance values between the media programs.
19. The non-transitory computer-readable storage medium of claim 14 , wherein combining comprises using a non-linear combination of:
F
(
x
)
=
lim
t
→
∞
F
t
(
x
)
=
∑
k
=
1
t
β
k
f
k
(
x
)
,
where F t (x) is a scoring function in the t-th iteration with respect to the single relevance table and F(x) is a final scoring function with respect to the single relevance table and β k is the weight of F k(x) .
20. An apparatus comprising:
one or more computer processors; and
a non-transitory computer-readable storage medium comprising instructions, that when executed, control the one or more computer processors to be configured for:
generating a plurality of sub-relevance tables including a first set of relevance values between a first set of media programs and a second set of media programs being provided by a video delivery service, wherein each table models relevance values relating similarity for a single feature in a plurality of features between the first set of media programs and the second set of media programs;
receiving labeling results including a second set of relevance values between the first set of media programs and the second set of media programs, the labeling results based on user input;
generating weights for each of the plurality of sub-relevance tables based on the second set of relevance values for the labeling results and the first set of relevance values of the plurality of sub-relevance tables;
combining relevance values from the plurality of sub-relevance tables into a single relevance table by applying respective weights for each of the plurality of sub-relevance tables to the relevance values, wherein the single relevance table includes a third set of relevance values relating similarity for the plurality of features between the first set of media programs and the second set of media programs; and
providing a recommendation to a user using the video delivery service, the recommendation generated using the third set of relevance values from the single relevance table and a characteristic of the user.