System, method, and non-transitory computer readable storage medium, for diversifying complementary item recommendations based on user preferences
A method including determining, in real-time, a diversity preference score for a user based at least in part on an anchor item chosen by the user via a user interface executed on a user device of the user. The method also can include determining, in real-time, a comparison result between the diversity preference score and a diversity preference threshold. The method further can include generating, in real-time, a personalized recommendation pool based on (a) the comparison result, (b) a complementary recommendation pool generated based at least in part on the anchor item, and (c) a diversity objective function. In many embodiments, when the comparison result indicates that the diversity preference score is greater than the diversity preference threshold, the diversity objective function can be associated with cross-domain diversity. In a number of embodiments, when the comparison result indicates that the diversity preference score is not greater than the diversity preference threshold, the diversity objective function can be associated with within-domain diversity. The method additionally can include transmitting, in real-time through the computer network, the personalized recommendation pool to be displayed with the anchor item on the user interface. Other embodiments are disclosed.
1. A system comprising:
one or more processors; and
one or more non-transitory computer-readable media storing computing instructions that, when run on the one or more processors, cause the one or more processors to perform:
training iteratively an embedding-based machine learning module to generate a recommendation pool for a reference item, comprising:
converting each item attribute of item attributes into a respective predefined data format of one or more predefined attribute formats;
generating item embeddings based on the item attributes, as converted, for training items;
generating customer embeddings based on customer behavior data for customers;
training the embedding-based machine learning module to determine customer item preference coefficients based on the item embeddings and the customer embeddings;
training the embedding-based machine learning module to determine a respective likelihood of co-purchase between a first item and a second item of the training items for each customer of the customers based on the customer item preference coefficients; and
re-training the embedding-based machine learning module further based at least in part on historic behavior of the embedding-based machine learning module;
determining, in real-time, a diversity preference score for a user based at least in part on: (a) an anchor item chosen by the user via a user interface executed on a user device of the user and (b) a department count for items previously purchased by the user with or after the anchor item, comprising:
determining a respective department for each of a predetermined number of items previously purchased by the user after the user purchased an item in an anchor department of the anchor item; and
determining, in real-time, a count of different departments among the respective departments for the predetermined number of items, wherein:
determining, in real-time, the diversity preference score further comprises determining, in real-time, the diversity preference score based further at least in part on the count of different departments;
determining, in real-time, a comparison result between the diversity preference score and a diversity preference threshold;
generating, in real-time, a personalized recommendation pool based on (a) the comparison result, (b) a complementary recommendation pool generated by the embedding-based machine learning module based at least in part on the anchor item, and (c) a diversity objective function, wherein:
generating the personalized recommendation pool comprises re-ranking recommended items of the complementary recommendation pool based on the diversity objective function and, when the diversity preference score is greater than the diversity preference threshold, the re-ranking includes adding a recommended item one at a time based on a fast greedy maximum a posteriori (MAP) approach and a cross-domain diversity objective function;
the recommended items of the complementary recommendation pool are ranked according to a respective personalized complementary score for each of the recommended items of the complementary recommendation pool;
when the comparison result indicates that the diversity preference score is greater than the diversity preference threshold, the diversity objective function is associated with cross-department diversity for increasing a recommended-item-department count for the personalized recommendation pool; and
when the comparison result indicates that the diversity preference score is not greater than the diversity preference threshold, the diversity objective function is associated with within-department diversity for increasing an anchor-department-item count for the personalized recommendation pool associated with the anchor department of the anchor item; and
transmitting, in real-time, the personalized recommendation pool to be displayed with the anchor item on the user interface.
2. The system in claim 1 , wherein the computing instructions, when run on the one or more processors, further cause the one or more processors to perform:
reducing a size of the complementary recommendation pool based on a predetermined size limit.
3. The system in claim 1 , wherein:
generating the personalized recommendation pool further comprises:
adding a recommended item from the recommended items of the complementary recommendation pool to the personalized recommendation pool iteratively based at least in part on a determinantal point process (DPP) with fast greedy maximum a posteriori (MAP) inference.
4. The system in claim 1 , wherein the diversity preference threshold is set by the embedding-based machine learning module, as trained.
5. The system in claim 2 , wherein the predetermined size limit is determined based on a type of user interface for displaying the personalized recommendation pool.
6. The system in claim 3 , wherein:
adding the recommended item from the recommended items of the complementary recommendation pool to the personalized recommendation pool iteratively further comprises
determining a respective diversity score for each of one or more remaining items of the recommended items of the complementary recommendation pool.
7. The system in claim 6 , wherein:
when the diversity objective function is associated with the cross-department diversity, the respective diversity score for each of the one or more remaining items of the recommended items of the complementary recommendation pool is calculated based in part on: (a) a respective personalized complementary score for the each of the one or more remaining items; and (b) a diversification increment by the each of the one or more remaining items; and
when the diversity objective function is associated with the within-department diversity, the respective diversity score for each of the one or more remaining items of the recommended items of the complementary recommendation pool is calculated based in part on: (a) the respective personalized complementary score for the each of the one or more remaining items; and (b) a similarity increment by the each of the one or more remaining items.
8. The system in claim 6 , wherein:
when the diversity objective function is associated with the cross-department diversity, the respective diversity score for the each of the one or more remaining items is calculated by:
(α S q,r +(1−α)(log(det( L R d +[r] ))−log(det( L R d )))], wherein
S q,r : a respective personalized complementary score between the anchor item (q) and the each of the one or more remaining items (r) for the user,
S
q
,
r
=
1
+
∫
q
T
g
r
2
,
and f q T and g r is a product score for normalized item embeddings between q and r;
α: a predetermined constant, 0≤α<1;
det(L R ): a determinant of a kernel matrix L indexed by member items of an item pool (R), R comprising a subset of the complementary recommendation pool; and
L R : an item-to-item similarity matrix for the member items in the item pool
(
R
)
,
L
R
=
1
+
H
T
H
2
,
H
≡
{
g
r
|
r
∈
R
}
.
9. The system in claim 6 , wherein:
when the diversity objective function is associated with the within-department diversity, the respective diversity score for the each of the one or more remaining items is calculated by:
(β S q,r +(1−β)(log(det( L′ R s +[r] ))−log(det( L′ R s )))], wherein:
S q,r : a respective personalized complementary score between the anchor item (g) and the each of the one or more remaining items (r) for the user,
S
q
,
r
=
1
+
∫
q
T
g
r
2
,
and f q T and g r is a product score for normalized item embeddings between q and r;
β: a predetermined constant, 0≤“β”<1;
det(L′ R ): a determinant of a kernel matrix L′indexed by member items of an item pool (R), R comprising a subset of the complementary recommendation pool and extra complementary items; and
L′ R : an item-to-item dissimilarity matrix for the member items in the item pool (R), L′=1+diag(L)−L, and
L
R
=
1
+
H
T
H
2
,
H≡{g r |r∈R}.
10. The system in claim 6 , wherein:
adding the recommended item from the recommended items of the complementary recommendation pool to the personalized recommendation pool iteratively further comprises adding the recommended item of the one or more remaining items to the personalized recommendation pool when the recommended item is associated with a maximal respective diversity score of the respective diversity score for each of the one or more remaining items.
11. A method being implemented via execution of computing instructions configured to run at one or more processors and stored at one or more non-transitory computer-readable media, the method comprising:
training iteratively an embedding-based machine learning module to generate a recommendation pool for a reference item, comprising:
converting each item attribute of item attributes into a respective predefined data format of one or more predefined attribute formats;
generating item embeddings based on the item attributes, as converted, for training items;
generating customer embeddings based on customer behavior data for customers;
training the embedding-based machine learning module to determine customer item preference coefficients based on the item embeddings and the customer embeddings;
training the embedding-based machine learning module to determine a respective likelihood of co-purchase between a first item and a second item of the training items for each customer of the customers based on the customer item preference coefficients; and
re-training the embedding-based machine learning module based at least in part on historic behavior of the embedding-based machine learning module;
determining, in real-time, a diversity preference score for a user based at least in part on:
(a) an anchor item chosen by the user via a user interface executed on a user device of the user and (b) a department count for items previously purchased by the user with or after the anchor item, comprising:
determining a respective department for each of a predetermined number of items previously purchased by the user after the user purchased an item in an anchor department of the anchor item; and
determining, in real-time, a count of different departments among the respective departments for the predetermined number of items, wherein:
determining, in real-time, the diversity preference score further comprises determining, in real-time, the diversity preference score based further at least in part on the count of different departments;
determining, in real-time, a comparison result between the diversity preference score and a diversity preference threshold;
generating, in real-time, a personalized recommendation pool based on (a) the comparison result, (b) a complementary recommendation pool generated by the embedding-based machine learning module based at least in part on the anchor item, and (c) a diversity objective function, wherein:
generating the personalized recommendation pool comprises re-ranking recommended items of the complementary recommendation pool based on the diversity objective function and, when the diversity preference score is greater than the diversity preference threshold, the re-ranking includes adding a recommended item one at a time based on a fast greedy maximum a posteriori (MAP) approach and a cross-domain diversity objective function;
the recommended items of the complementary recommendation pool are ranked according to a respective personalized complementary score for each of the recommended items of the complementary recommendation pool;
when the comparison result indicates that the diversity preference score is greater than the diversity preference threshold, the diversity objective function is associated with cross-department diversity for increasing a recommended-item-department count for the personalized recommendation pool; and
when the comparison result indicates that the diversity preference score is not greater than the diversity preference threshold, the diversity objective function is associated with within-department diversity for increasing an anchor-department-item count for the personalized recommendation pool associated with the anchor department of the anchor item; and
transmitting, in real-time, the personalized recommendation pool to be displayed with the anchor item on the user interface.
12. The method in claim 11 further comprising:
reducing a size of the complementary recommendation pool based on a predetermined size limit.
13. The method in claim 11 , wherein:
generating the personalized recommendation pool further comprises:
adding a recommended item from the recommended items of the complementary recommendation pool to the personalized recommendation pool iteratively based at least in part on a determinantal point process (DPP) with fast greedy maximum a posteriori (MAP) inference.
14. The method in claim 11 , wherein the diversity preference threshold is set by the embedding-based machine learning module, as trained.
15. The method in claim 12 , wherein the predetermined size limit is determined based on a type of user interface for displaying the personalized recommendation pool.
16. The method in claim 13 , wherein:
adding the recommended item of the recommended items of the complementary recommendation pool to the personalized recommendation pool iteratively further comprises:
determining a respective diversity score for each of one or more remaining items of the recommended items of the complementary recommendation pool; and
adding the recommended item of the one or more remaining items to the personalized recommendation pool when the recommended item is associated with a maximal respective diversity score of the respective diversity score for each of the one or more remaining items.
17. The method in claim 16 , wherein:
when the diversity objective function is associated with the cross-department diversity, the respective diversity score for each of the one or more remaining items of the recommended items of the complementary recommendation pool is calculated based in part on: (a) a respective personalized complementary score for the each of the one or more remaining items; and (b) a diversification increment by the each of the one or more remaining items; and
when the diversity objective function is associated with the within-department diversity, the respective diversity score for each of the one or more remaining items of the recommended items of the complementary recommendation pool is calculated based in part on: (a) the respective personalized complementary score for the each of the one or more remaining items; and (b) a similarity increment by the each of the one or more remaining items.
18. The method in claim 16 , wherein:
when the diversity objective function is associated with the cross-department diversity, the respective diversity score for the each of the one or more remaining items is calculated by:
(α S q,r +(1−α)(log(det( L R d +[r] ))−log(det( L R d )))], wherein:
S q,r : a respective personalized complementary score between the anchor item (q) and the each of the one or more remaining items (r) for the user,
S
q
,
r
=
1
+
∫
q
T
g
r
2
,
and f q T and g r is a product score for normalized item embeddings between q and r;
α: a predetermined constant, 0≤α<1;
det(L R ): a determinant of a kernel matrix L indexed by member items of an item pool (R), R comprising a subset of the complementary recommendation pool; and
L R : an item-to-item similarity matrix for the member items in the item pool
(
R
)
,
L
R
=
1
+
H
T
H
2
,
H≡{g r |r∈R}.
19. The method in claim 16 , wherein:
when the diversity objective function is associated with the within-department diversity, the respective diversity score for the each of the one or more remaining items is calculated by:
(β S q,r +(1−β)(log(det( L′ R s +[r] ))−log(det( L′ R s )))], wherein:
S q,r : a respective personalized complementary score between the anchor item (q) and the each of the one or more remaining items (r) for the user,
S
q
,
r
=
1
+
∫
q
T
g
r
2
,
and f q T and g r is a product score for normalized item embeddings between q and r;
β: a predetermined constant, 0≤“β”<1;
det(L′ R ): a determinant of a kernel matrix L′indexed by member items of an item pool (R),f R comprising a subset of the complementary recommendation pool and extra complementary items; and
L′ R : an item-to-item dissimilarity matrix for the member items in the item pool (R), L′=1+diag(L)−L, and
L
R
=
1
+
H
T
H
2
,
≡{g r |r∈R}.
20. A non-transitory computer readable storage medium storing one or more computing instructions that, when run on one or more processors, cause the one or more processors to perform operations comprising:
training iteratively an embedding-based machine learning module to generate a recommendation pool for a reference item, comprising:
converting each item attribute of item attributes into a respective predefined data format of one or more predefined attribute formats;
generating item embeddings based on the item attributes, as converted, for training items;
generating customer embeddings based on customer behavior data for customers;
training the embedding-based machine learning module to determine customer item preference coefficients based on the item embeddings and the customer embeddings;
training the embedding-based machine learning module to determine a respective likelihood of co-purchase between a first item and a second item of the training items for each customer of the customers based on the customer item preference coefficients; and
re-training the embedding-based machine learning module based at least in part on historic behavior of the embedding-based machine learning module;
determining, in real-time, a diversity preference score for a user based at least in part on:
(a) an anchor item chosen by the user via a user interface executed on a user device of the user and (b) a department count for items previously purchased by the user with or after the anchor item, comprising:
determining a respective department for each of a predetermined number of items previously purchased by the user after the user purchased an item in an anchor department of the anchor item; and
determining, in real-time, a count of different departments among the respective departments for the predetermined number of items, wherein:
determining, in real-time, the diversity preference score further comprises determining, in real-time, the diversity preference score based further at least in part on the count of different departments;
determining, in real-time, a comparison result between the diversity preference score and a diversity preference threshold;
generating, in real-time, a personalized recommendation pool based on (a) the comparison result, (b) a complementary recommendation pool generated by the embedding-based machine learning module based at least in part on the anchor item, and (c) a diversity objective function, wherein:
generating the personalized recommendation pool comprises:
generating the personalized recommendation pool comprises re-ranking recommended items of the complementary recommendation pool based on the diversity objective function and, when the diversity preference score is greater than the diversity preference threshold, the re-ranking includes adding a recommended item one at a time based on a fast greedy maximum a posteriori (MAP) approach and a cross-domain diversity objective function;
the recommended items of the complementary recommendation pool are ranked according to a respective personalized complementary score for each of the recommended items of the complementary recommendation pool;
when the comparison result indicates that the diversity preference score is greater than the diversity preference threshold, the diversity objective function is associated with cross-department diversity for increasing a recommended-item-department count for the personalized recommendation pool; and
when the comparison result indicates that the diversity preference score is not greater than the diversity preference threshold, the diversity objective function is associated with within-department diversity for increasing an anchor-department-item count for the personalized recommendation pool associated with the anchor department of the anchor item; and
transmitting, in real-time, the personalized recommendation pool to be displayed with the anchor item on the user interface.