Content recommendation
Briefly, embodiments of methods and/or systems for performing content recommendation are disclosed. For one embodiment, as an example, estimating relevance may include computing an inner product of latent factors corresponding to a plurality of users and features of one or more content items.
1. A computer-implemented method, comprising:
obtaining a plurality of latent factors for a plurality of users based on positive interactions and negative interactions by the plurality of users with a plurality of content items, a first negative interaction of the negative interactions indicating that a user has skipped an opportunity to select or click on at least one content item of the plurality of content items and that the user is disinterested in the at least one content item of the plurality of content items, a second negative interaction of the negative interactions indicating that the user spent less than a threshold period of time viewing a second content item, of the plurality of content items, comprising an article;
generating a plurality of latent factor vectors using the plurality of latent factors, including latent factors corresponding to the negative interactions, for the plurality of users, each of the plurality of latent factor vectors corresponding to a different one of the plurality of users;
generating a first latent factor matrix based, at least in part, on the plurality of latent factor vectors, including the latent factors corresponding to the negative interactions, and using a first stochastic gradient descent;
generating a plurality of content item features for the plurality of content items, each content item being represented by a corresponding subset of the plurality of content item features;
generating a second latent factor matrix using a second stochastic gradient descent, the second latent factor matrix mapping the plurality of content item features for the plurality of content items to the plurality of latent factors, including the latent factors corresponding to the negative interactions, for the plurality of users;
generating a user profile matrix based, at least in part, on the first latent factor matrix and the second latent factor matrix, wherein a first user profile, of a first user, associated with the user profile matrix comprises (i) a first component comprising a common mapping, shared between multiple users including the first user, from content item features to latent factors and (ii) a second component comprising latent factors of the first user, wherein a first feature is included in the first user profile responsive to the first user consuming content associated with the first feature at a consumption rate exceeding a threshold rate;
estimating relevance of a first content item to the first user based on an inner product of two or more latent factors, including at least one of the latent factors corresponding to the negative interactions, of the user profile matrix corresponding to the first user and the subset of the plurality of content item features corresponding to the first content item, wherein the two or more latent factors of the user profile matrix comprise a latent factor indicative of an interest of the first user in a first topic based on a determination that the first user interacted with one or more websites associated with a second topic and one or more other users that interacted with one or more websites associated with the second topic also interacted with one or more websites associated with the first topic; and
generating a recommendation of the first content item for the first user based on the relevance.
2. The method of claim 1 , at least one of the positive interactions corresponding to a user click.
3. The method of claim 1 , wherein the plurality of content item features comprises at least one of named entities, topical categories, noun-phrase tokens, or n-grams.
4. The method of claim 1 , wherein the positive interactions are weighted more heavily than the negative interactions.
5. The method of claim 1 , wherein obtaining the plurality of latent factors comprises:
applying a feedback matrix including a plurality of values, each of a first subset of the plurality of values representing a corresponding one of the positive interactions and each of a second subset of the plurality of values representing a corresponding one of the negative interactions.
6. The method of claim 5 , each of the first subset of the plurality of values including a first value and each of the second subset of the plurality of values including a second value.
7. The method of claim 6 , the first value being a positive value and the second value being a negative value.
8. The method of claim 1 , each of the positive interactions being represented by a positive value and each of the negative interactions being represented by a negative value.
9. The method of claim 1 , wherein at least one of generating the first latent factor matrix or generating the second latent factor matrix comprises applying at least one stochastic gradient descent until the first latent factor matrix and the second latent factor matrix converge.
10. The method of claim 1 , wherein at least one of generating the first latent factor matrix or generating the second latent factor matrix comprises:
computing the first latent factor matrix and the second latent factor matrix according to a square loss approach.
11. An apparatus comprising:
one or more processors and memory configured to:
obtain a plurality of latent factors for a plurality of users based on positive interactions and negative interactions by the plurality of users with a plurality of content items, a first negative interaction of the negative interactions indicating that a user has skipped an opportunity to select or click on at least one content item of the plurality of content items and that the user is disinterested in the at least one content item of the plurality of content items, a second negative interaction of the negative interactions indicating that the user spent less than a threshold period of time viewing a second content item, of the plurality of content items, comprising an article;
generate a plurality of latent factor vectors using the plurality of latent factors, including latent factors corresponding to the negative interactions, for the plurality of users, each of the plurality of latent factor vectors corresponding to a different one of the plurality of users;
generate a first latent factor matrix based, at least in part, on the plurality of latent factor vectors, including the latent factors corresponding to the negative interactions, and using a first stochastic gradient descent;
generate a plurality of content item features for the plurality of content items, each content item being represented by a corresponding subset of the plurality of content item features;
generate a second latent factor matrix using a second stochastic gradient descent, the second latent factor matrix mapping the plurality of content item features for the plurality of content items to the plurality of latent factors, including the latent factors corresponding to the negative interactions, for the plurality of users;
generate a user profile matrix based, at least in part, on the first latent factor matrix and the second latent factor matrix, wherein a first user profile, of a first user, associated with the user profile matrix comprises (i) a first component comprising a common mapping, shared between multiple users including the first user, from content item features to latent factors and (ii) a second component comprising latent factors of the first user;
estimate relevance of a first content item to the first user based on an inner product of two or more latent factors, including at least one of the latent factors corresponding to the negative interactions, of the user profile matrix corresponding to the first user and the subset of the plurality of content item features corresponding to the first content item, wherein the two or more latent factors of the user profile matrix comprise a latent factor indicative of an interest of the first user in a first topic based on a determination that the first user interacted with one or more websites associated with a second topic and one or more other users that interacted with one or more websites associated with the second topic also interacted with one or more websites associated with the first topic; and
generate a recommendation of the first content item for the first user based on the relevance.
12. The apparatus of claim 11 , wherein the positive interactions are weighted more heavily than the negative interactions.
13. The apparatus of claim 11 , the one or more processors and memory being configured to:
obtain the plurality of latent factors based, at least in part, on a feedback matrix including a plurality of values, each of a first subset of the plurality of values representing a corresponding one of the positive interactions and each of a second subset of the plurality of values representing a corresponding one of the negative interactions.
14. The apparatus of claim 11 , each of the positive interactions being represented by a positive value and each of the negative interactions being represented by a negative value.
15. The apparatus of claim 11 , the one or more processors and memory being configured to:
generate at least one of the first latent factor matrix or the second latent factor matrix by applying at least one stochastic gradient descent until the first latent factor matrix and the second latent factor matrix converge.
16. The apparatus of claim 11 , the one or more processors and memory being configured to:
generate at least one of the first latent factor matrix or the second latent factor matrix using at least one stochastic gradient descent by computing the first latent factor matrix and the second latent factor matrix according to a square loss approach.
17. An apparatus comprising:
one or more processors; and
memory comprising instructions that when executed by at least one of the one or more processors perform operations comprising:
obtaining a plurality of latent factors for a plurality of users based on positive interactions and negative interactions by the plurality of users with a plurality of content items, a first negative interaction of the negative interactions indicating that a user has skipped an opportunity to select or click on at least one content item of the plurality of content items and that the user is disinterested in the at least one content item of the plurality of content items, a second negative interaction of the negative interactions indicating that the user spent less than a threshold period of time viewing a second content item, of the plurality of content items, comprising an article;
generating a plurality of latent factor vectors using the plurality of latent factors, including latent factors corresponding to the negative interactions, for the plurality of users, each of the plurality of latent factor vectors corresponding to a different one of the plurality of users;
generating a first latent factor matrix based, at least in part, on the plurality of latent factor vectors, including the latent factors corresponding to the negative interactions, and using a first stochastic gradient descent;
generating a plurality of content item features for the plurality of content items, each content item being represented by a corresponding subset of the plurality of content item features;
generating a second latent factor matrix using a second stochastic gradient descent, the second latent factor matrix mapping the plurality of content item features for the plurality of content items to the plurality of latent factors, including the latent factors corresponding to the negative interactions, for the plurality of users;
generating a user profile matrix based, at least in part, on the first latent factor matrix and the second latent factor matrix, wherein a first user profile, of a first user, associated with the user profile matrix comprises (i) a first component comprising a common mapping, shared between multiple users including the first user, from content item features to latent factors and (ii) a second component comprising latent factors of the first user;
estimating relevance of a first content item to the first user based on an inner product of two or more latent factors, including at least one of the latent factors corresponding to the negative interactions, of the user profile matrix corresponding to the first user and the subset of the plurality of content item features corresponding to the first content item, wherein the two or more latent factors of the user profile matrix comprise a latent factor indicative of an interest of the first user in a first topic based on a determination that the first user interacted with one or more websites associated with a second topic and one or more other users that interacted with one or more websites associated with the second topic also interacted with one or more websites associated with the first topic; and
generating a recommendation of the first content item for the first user based on the relevance.
18. The apparatus of claim 17 , wherein the positive interactions are weighted more heavily than the negative interactions.
19. The apparatus of claim 17 , each of the positive interactions being represented by a positive value and each of the negative interactions being represented by a negative value.