Systems and methods of providing recommendations of content items
A method of recommending content items includes obtaining vector representations of items based on a matrix of items versus item users that indicates item use by the users, reducing each of the vector representations to a two-dimensional space, creating clusters having cluster centers using one or more mixture models based on the reduced vector representations, using the cluster centers for one or more mixture models to create discrete categories to which items can be assigned and providing one or more recommendations to a first user based on the item assignments within the discrete categories.
1. A method of automatically generating item recommendations for users in a system having a processor, and a database, and connected to a user device through a network, the method comprising:
obtaining a matrix for a plurality of users and a plurality of items, the matrix indicating whether each user has previously consumed each item;
generating, using the processor, an n-dimensional vector representation for each user and an m-dimensional vector representation for each item, wherein each of n and m is at least 2;
reducing, using the processor, each user vector representation and each item vector representation to a two-dimensional space;
creating a plurality of clusters having cluster centers by using one or more first mixture models based on the reduced vector representations, the one or more first mixture models executed via the processor to create the plurality of clusters;
using the cluster centers for one or more second mixture models to create discrete categories to which each of the plurality of items are assigned, the one or more second mixture models executing via the processor to create the discrete categories and to determine a probability of each of the plurality of items being assigned to one or more of each of the discrete categories, wherein at least one of the items is assigned to multiple categories;
identifying an item of interest from the plurality of items assigned to the discrete categories that include at least one item previously consumed; and
providing a recommendation of the item of interest to the user device via the network.
2. The method of claim 1 , wherein generating the vectors comprises using collaborative filtering (CF) to construct a user vector {right arrow over (v u )} and an item vector {right arrow over (v l )}, for all users u and items i respectively.
3. The method of claim 2 , wherein the recommendation is further based on a probability p u,i of a user u consuming an item i determined as:
p
u
,
i
=
exp
(
v
→
u
·
v
→
i
)
exp
(
v
→
u
·
v
→
i
)
+
1
.
4. The method of claim 1 , wherein the second one or more mixture models includes at least one Gaussian mixture model.
5. The method of claim 1 , wherein the vector representations are reduced to a two-dimensional space using t-distributed stochastic neighbor embedding (t-SNE).
6. The method of claim 5 , wherein a probability score for a particular item belonging to a particular category i is factored based on a weight calculation as follows:
weight
i
=
exp
(
-
(
x
→
-
μ
ι
→
)
2
2
σ
i
2
)
where x is a location of the t-SNE projected particular item vector, μ i is a center or mean vector of category i, and σ i is a standard deviation of cluster of category i.
7. The method of claim 1 , wherein the one or more first mixture models uses a nonparametric process that does not require a predefined number of clusters.
8. The method of claim 7 , wherein the one or more first mixture models include at least one Dirichlet process mixture models.
9. The method of claim 1 , wherein at least some of the plurality of items correspond to musician identifiers.
10. The method of claim 1 , wherein at least some of the plurality of items correspond to album identifiers.
11. The method of claim 1 , wherein at least some of the plurality of items correspond to book identifiers.
12. The method of claim 1 , wherein the recommendation is provided to the user device as part of a streaming music service.
13. The method of claim 1 , wherein collaborative filtering matrix is based on data obtained about the user device from a third party service.
14. A system configured to automatically generate item recommendations for users, the system comprising:
a memory configured to store instructions;
a network connecting the system to a user device; and
a processor configured to execute the instructions to:
obtain a matrix for a plurality of users and a plurality of items, the matrix indicating whether each user has previously consumed each item;
generate an n-dimensional vector representation for each user and an m-dimensional vector representation for each item, wherein each of n and m is at least 2;
reduce each user vector representation and each item vector representation to a two-dimensional space;
create a plurality of clusters having cluster centers by using one or more first mixture models based on the reduced vector representations;
use the cluster centers for one or more second mixture models to create discrete categories to which each of the plurality of items are assigned;
determine a probability, using the second one or more mixture models, of each of the plurality of items being assigned to each of the discrete categories, wherein at least one of the items is assigned to multiple categories;
identify an item of interest from the plurality of items assigned to the discrete categories that include at least one item previously consumed; and
provide a recommendation of the item of interest to the user device via the network.
15. The system of claim 14 , wherein the processor is configured to generate the vectors using collaborative filtering (CF) to construct a user vector {right arrow over (v u )} and an item vector {right arrow over (v l )}, for all users u and items i respectively.
16. The system of claim 14 , wherein the processor is further configured to provide the one or more recommendations based on a probability p u,i of a user u consuming an item i determined as:
p
u
,
i
=
exp
(
v
→
u
·
v
→
i
)
exp
(
v
→
u
·
v
→
i
)
+
1
.
17. The system of claim 14 , wherein the second one or more mixture models includes at least one Gaussian mixture model.
18. The system of claim 14 , wherein the processor is further configured reduce the vector representations to a two-dimensional space using t-distributed stochastic neighbor embedding (t-SNE).
19. The system of claim 18 , wherein the processor is configured to factor a probability score for a particular item belonging to a particular category i based on a weight calculation as follows:
weight
i
=
exp
(
-
(
x
→
-
μ
ι
→
)
2
2
σ
i
2
)
where x is a location of the t-SNE projected particular item vector, μ i is a center or mean vector of category i, and σ i is a standard deviation of cluster of category i.
20. The system of claim 14 , wherein the one or more first mixture models uses a nonparametric process that does not require a predefined number of clusters.
21. The system of claim 20 , wherein the one or more first mixture models include at least one Dirichlet process mixture models.
22. The system of claim 14 , wherein at least some of the plurality of items correspond to musician identifiers.
23. The system of claim 14 , wherein at least some of the plurality of items correspond to album identifiers.
24. The system of claim 14 , wherein at least some of the plurality of items correspond to book identifiers.
25. The system of claim 14 , wherein the processor is configured provide the recommendation to the user device as part of a streaming music service.
26. The system of claim 14 , wherein collaborative filtering matrix is based on data obtained about the user device from a third party service.