System, non-transitory computer readable medium, and method for personalized complementary recommendations

1. A system, comprising:

a non-transitory memory having instructions stored thereon and a processor configured to read the instructions to:

generate, by an iterative training process using a first training data set including add-to-cart data, a trained supervised model;

generate, by an iterative training process using a second training data set including a triple including a user, an anchor item, and a recommendation item, a trained unsupervised model, wherein the trained unsupervised model is configured to minimize a loss function between at least two vectors in a triple of u,i,j , where u is a vector representation of a user, i is a vector representation of an anchor item, and j is a vector representation of an item co-purchased with the anchor item;

receive a user identifier and an anchor item identifier;

generate, by the trained supervised model, a set of personalized-weighted items based on the user identifier and the anchor item identifier;

generate, by the trained unsupervised model, a set of complementary-weighted items based on the user identifier; and

generate a set of personalized complementary recommendations including a subset of the personalized-weighted items and a subset of the complementary-weighted items.

2. The system of claim 1 , wherein the trained supervised model applies a logistic regression to a predetermined set of features to generate the set of personalized-weighted items.

3. The system of claim 2 , wherein the predetermined set of features includes a feature f where:

f u,i,j =Σ u m 1/ r

where m is a number of add-to-cart sequences in which a user u adds an item i to a cart and adds an item j to the cart within a time window k, and r is the position of item j when item j is added to the cart.

4. The system of claim 2 , wherein the predetermined set of features includes a feature representative of a recency of a recommended item bought by a user.

5. The system of claim 1 , wherein the loss function is represented as:

= u,i + u,j + i,j ,u∈U,i∈V a ,j∈V r

where U is vector space of the user, V a , is a vector space of the anchor item, and V r is a vector space of the co-purchased item.

6. The system of claim 1 , wherein the processor is configured to read the instructions to generate the set of personalized complementary recommendations to:

generate a set of ranked complimentary-weighted items by ranking the set of complimentary-weighted items according to at least one metric;

generate a set of combined items including at least one of the personalized-weighted items and at least one of the ranked complimentary-weighted items; and

select k top-ranked items from the set of combined items.

7. The system of claim 6 , wherein the at least one metric comprises a distance between an anchor item i and a recommended item j.

8. A non-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by a processor cause a device to perform operations comprising:

generating, by an iterative training process using a first training data set including add-to-cart data, a trained supervised model;

generating, by an iterative training process using a second training data set including a triple including a user, an anchor item, and a recommendation item, a trained unsupervised model, wherein the trained unsupervised model is configured to minimize a loss function between at least two vectors in a triple of u,i,j , where u is a vector representation of a user, i is a vector representation of an anchor item, and j is a vector representation of an item co-purchased with the anchor item;

receiving a user identifier and an anchor item identifier;

generating, by the trained supervised model, a set of personalized-weighted items based on the user identifier and the anchor item identifier;

generating, by the trained unsupervised model, a set of complementary-weighted items based on the user identifier and the anchor item identifier; and

generating a set of personalized complementary recommendations including a subset of the personalized-weighted items and a subset of the complementary-weighted items.

9. The non-transitory computer readable medium of claim 8 , wherein the trained supervised model applies a logistic regression to a predetermined set of features to generate the set of personalized-weighted items.

10. The non-transitory computer readable medium of claim 9 , wherein the predetermined set of features includes a feature f where:

f u,i,j =Σ u m 1/ r

where m is a number of add-to-cart sequences in which a user u adds an item i to a cart and adds an item j to the cart within a time window k, and r is the position of item j when item j is added to the cart.

11. The non-transitory computer readable medium of claim 9 , wherein the predetermined set of features includes a feature representative of a recency of a recommended item bought by a user.

12. The non-transitory computer readable medium of claim 8 , wherein the loss function is represented as:

= u,i + u,j + i,j ,u∈U,i∈V a ,j∈V r

where U is vector space of the user, V a , is a vector space of the anchor item, and V r is a vector space of the co-purchased item.

13. The non-transitory computer readable medium of claim 8 , wherein generating the set of personalized complementary recommendations comprises:

generating a set of ranked complimentary-weighted items by ranking the set of complimentary-weighted items according to at least one metric;

generating a set of combined items including at least one of the personalized-weighted items and at least one of the ranked complementary-weighted items; and

selecting k top-ranked items from the set of combined items.

14. The non-transitory computer readable medium of claim 13 , wherein the at least one metric comprises a distance between an anchor item i and a recommended item j.

15. A computer-implemented method, comprising:

generating, by an iterative training process using a first training data set including add-to-cart data, a trained supervised model;

generating, by an iterative training process using a second training data set including a triple including a user, an anchor item, and a recommendation item, a trained unsupervised model, wherein the trained unsupervised model is configured to minimize a loss function between at least two vectors in a triple of u,i,j , where u is a vector representation of a user, i is a vector representation of an anchor item and j is a vector representation of an item co-purchased with the anchor item, and wherein the loss function is represented as:

= u,i + u,j + i,j ,u∈U,i∈V a ,j∈V r

where U is vector space of the user, V a is a vector space of the anchor item, and V r is a vector space of the co-purchased item; and

receiving a user identifier and an anchor item identifier;

generating a set of personalized-weighted items, by the trained supervised model, based on the user identifier and the anchor item identifier;

generating a set of complementary-weighted items, by the trained unsupervised model, based on the user identifier and the anchor item identifier; and

generating a set of personalized complementary recommendations including a subset of the personalized-weighted items and a subset of the complementary-weighted items.

16. The computer-implemented method of claim 15 , wherein the trained supervised model applies a logistic regression to a predetermined set of features to generate the set of personalized-weighted items, wherein the predetermined set of features includes a feature f where:

f u,i,j =Σ u m 1/ r

where m is a number of add-to-cart sequences in which a user u adds an item i to a cart and adds an item j to the cart within a time window k, and r is the position of item j when item j is added to the cart, and wherein the predetermined set of features includes a feature representative of a recency of a recommended item bought by a user.

17. The computer-implemented method of claim 15 , wherein generating the set of personalized complementary recommendations comprises:

generating a set of ranked complimentary-weighted items by ranking the set of complimentary-weighted items according a distance between an anchor item i and a recommended item j and a distance between a user u and the recommended item j;

generating a set of combined items including at least one of the personalized-weighted items and at least one of the ranked complementary-weighted items; and

selecting k top-ranked items from the set of combined items.