System and method for estimating user ratings from user behavior and providing recommendations
First Claim
Patent Images
1. A recommendation system for providing a recommendation set of recommended items from an item data store to a user, the recommendation system comprising:
- a memory storing the item data store;
an interface adapted to allow the user to browse an item set made from items in the item data store;
a processor connected to the memory and the interface, the processor comprising;
a ratings estimation module for;
receiving item property data comprising at least one item attribute for each item in the item set, wherein the at least one item attribute is a property of the item;
receiving training set data including user-specific information, wherein the user-specific information comprises at least one user-specific property that relates only to the user, and wherein the at least one user-specific property includes viewing time of at least one item browsed by the user, wherein the viewing time is an individual viewing time the user viewed the specific item; and
generating a training set of items comprising the at least one item browsed by the user; and
generating an item rating matrix by estimating user ratings for the items in the training set based on at least one rating predictor variable;
wherein each rating predictor variable comprises one of the at least one user-specific property and the at least one item attribute;
wherein an entry in the item rating matrix for a training set item e is generated according to;
Item Ratinge=f(x1, x2, . . . , xn)
wherein each xi is a value of the ith rating predictor variable, n is the number of rating predictor variables, and f is one of a non-linear combination and a linear combination of the rating predictor variables;
wherein at least one of the rating predictor variables is the viewing time; and
,a recommendation module connected to the ratings estimation module for receiving the item rating matrix and the item set, wherein the recommendation module (i) estimates user ratings for unseen items in the item set based on the item rating matrix without using information about any other user and (ii) generates the recommendation set based on the estimated ratings.
2 Assignments
0 Petitions
Accused Products
Abstract
A recommendation system and method for generating a set of recommended items for a user who browses items from an item data store. The recommendation system estimates ratings for at least some of the items previously viewed by the user, and then constructs a representation of the user'"'"'s implicit ratings of attributes of these items. These implicit attribute ratings are then used to estimate the user'"'"'s ratings for unseen items in the item data store.
-
Citations
34 Claims
-
1. A recommendation system for providing a recommendation set of recommended items from an item data store to a user, the recommendation system comprising:
-
a memory storing the item data store; an interface adapted to allow the user to browse an item set made from items in the item data store; a processor connected to the memory and the interface, the processor comprising; a ratings estimation module for; receiving item property data comprising at least one item attribute for each item in the item set, wherein the at least one item attribute is a property of the item; receiving training set data including user-specific information, wherein the user-specific information comprises at least one user-specific property that relates only to the user, and wherein the at least one user-specific property includes viewing time of at least one item browsed by the user, wherein the viewing time is an individual viewing time the user viewed the specific item; and generating a training set of items comprising the at least one item browsed by the user; and generating an item rating matrix by estimating user ratings for the items in the training set based on at least one rating predictor variable;
wherein each rating predictor variable comprises one of the at least one user-specific property and the at least one item attribute;
wherein an entry in the item rating matrix for a training set item e is generated according to;
Item Ratinge=f(x1, x2, . . . , xn)
wherein each xi is a value of the ith rating predictor variable, n is the number of rating predictor variables, and f is one of a non-linear combination and a linear combination of the rating predictor variables;
wherein at least one of the rating predictor variables is the viewing time; and
,a recommendation module connected to the ratings estimation module for receiving the item rating matrix and the item set, wherein the recommendation module (i) estimates user ratings for unseen items in the item set based on the item rating matrix without using information about any other user and (ii) generates the recommendation set based on the estimated ratings. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19)
-
-
20. A method for providing a recommendation set of recommended items from an item data store to a user who uses an interface to browse an item set made from items in the item data store, the method including:
-
a) obtaining item property data comprising at least one item attribute for each item in the item set, wherein the at least one item attribute is a property of the item, and training set data including user specific information, wherein the user specific information comprises at least one user-specific property that relates only to the user, and wherein the at least one user-specific property includes viewing time of at least one item browsed by the user, wherein the viewing time is an individual viewing time the user viewed the specific item; b) generating a training set of items comprising the at least one item browsed by the user; c) generating an item rating matrix by estimating user ratings for the items in the training set based on at least one rating predictor variable;
wherein each rating predictor variable comprises one of the at least one user-specific property and the at least one item attribute;
wherein an entry in the item rating matrix for a training set item e is generated according to;
Item Ratinge=f(x1, x2, . . . , xn)wherein each xi is a value of the ith rating predictor variable, n is the number of rating predictor variables, and f is one of a non-linear combination and a linear combination of the rating predictor variables;
wherein at least one of the rating predictor variables is the viewing time;d) estimating user ratings for unseen items in the item set based on the item rating matrix without using information about any other user; and
,e) generating the recommendation set based on the estimated ratings. - View Dependent Claims (21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34)
where c is a constant, each ai is a coefficient corresponding to a rating predictor variable, and each xi is a value of the ith rating predictor variable and where the coefficients ai are generated according to a model based on a combination of the at least one item attribute and the at least one user-specific property.
-
-
26. The method of claim 20, wherein estimating user ratings for unseen items includes:
-
a) generating pre-processed items by applying zero or more statistical transformation to the unseen items in the item set; b) generating a user profile having some subset of attributes of the pre-processed items along with at least one of an inferred user rating and explicit user rating, the inferred user rating being derived for some or all of the attributes from ratings in the item rating matrix; and
,c) generating the estimated user ratings for some or all of the unseen items in the item set based on the estimated user ratings for the items.
-
-
27. The method of claim 21, wherein the method further includes generating estimated user ratings for unseen items in the item set and using the estimated user ratings to generate the recommendation set, wherein estimating the user rating for an unseen item e is done according to
-
i = 1 R r i w i ∑ i = 1 W w i in which ri is a rating of the ith attribute, wi is relative importance of the ith attribute, |R| is the number of attributes in the user profile and |W| is the number weights.
-
-
28. The method of claim 26, wherein the method further includes characterizing the at least one item attribute by at least one of a text attribute and a numeric attribute.
-
29. The method of claim 28, wherein for estimating user ratings for unseen items in the item set, the method includes generating the numeric attribute in the user profile according to
-
j = 1 Tp r j q ij ∑ J = 1 Tp r j in which rj is the rating of the jth positive item in the training set, qij is the value of the ith numeric attribute for the jth positive item in the training set and |Tp| is the number of positive items and wherein a positive item is defined as an item having an attribute value greater than a predefined threshold value.
-
-
30. The method of claim 26, wherein for at least some numeric attributes, the method includes selecting a subset of items in the training set to generate the user profile, the subset being selected based on items in the training set having a rating greater than a predefined threshold value.
-
31. The method of claim 26, wherein the method includes generating the recommendation set by selecting n_rec unseen items having the highest estimated user ratings where n_rec is an integer value greater than or equal to one.
-
32. The method of claim 26, wherein the method includes generating the recommendation set by selecting the unseen items having an estimated user rating that is greater than a threshold value.
-
33. The method of claim 26 where the method includes placing a given unseen item in the recommendation set only when the estimated user rating of the given unseen item exceeds the ratings of all items previously viewed by the user.
-
34. The method of claim 20, wherein the method further includes processing the items in the item data store to generate the item property data.
Specification